]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
62f7fd21 MM |
2 | Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010 |
3 | Free Software Foundation, Inc. | |
b8698a0f | 4 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
5 | and Ira Rosen <irar@il.ibm.com> |
6 | ||
7 | This file is part of GCC. | |
8 | ||
9 | GCC is free software; you can redistribute it and/or modify it under | |
10 | the terms of the GNU General Public License as published by the Free | |
11 | Software Foundation; either version 3, or (at your option) any later | |
12 | version. | |
13 | ||
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
17 | for more details. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
20 | along with GCC; see the file COPYING3. If not see | |
21 | <http://www.gnu.org/licenses/>. */ | |
22 | ||
23 | #include "config.h" | |
24 | #include "system.h" | |
25 | #include "coretypes.h" | |
26 | #include "tm.h" | |
27 | #include "ggc.h" | |
28 | #include "tree.h" | |
29 | #include "target.h" | |
30 | #include "basic-block.h" | |
31 | #include "diagnostic.h" | |
cf835838 JM |
32 | #include "tree-pretty-print.h" |
33 | #include "gimple-pretty-print.h" | |
ebfd146a IR |
34 | #include "tree-flow.h" |
35 | #include "tree-dump.h" | |
36 | #include "cfgloop.h" | |
37 | #include "cfglayout.h" | |
38 | #include "expr.h" | |
39 | #include "recog.h" | |
40 | #include "optabs.h" | |
41 | #include "toplev.h" | |
42 | #include "tree-vectorizer.h" | |
43 | #include "langhooks.h" | |
44 | ||
45 | ||
46 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ | |
47 | ||
48 | /* Function vect_mark_relevant. | |
49 | ||
50 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
51 | ||
52 | static void | |
53 | vect_mark_relevant (VEC(gimple,heap) **worklist, gimple stmt, | |
54 | enum vect_relevant relevant, bool live_p) | |
55 | { | |
56 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
57 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
58 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
59 | ||
60 | if (vect_print_dump_info (REPORT_DETAILS)) | |
61 | fprintf (vect_dump, "mark relevant %d, live %d.", relevant, live_p); | |
62 | ||
63 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
64 | { | |
65 | gimple pattern_stmt; | |
66 | ||
b8698a0f | 67 | /* This is the last stmt in a sequence that was detected as a |
ebfd146a IR |
68 | pattern that can potentially be vectorized. Don't mark the stmt |
69 | as relevant/live because it's not going to be vectorized. | |
70 | Instead mark the pattern-stmt that replaces it. */ | |
71 | ||
72 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
73 | ||
74 | if (vect_print_dump_info (REPORT_DETAILS)) | |
75 | fprintf (vect_dump, "last stmt in pattern. don't mark relevant/live."); | |
76 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
77 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
78 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
79 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
80 | stmt = pattern_stmt; | |
81 | } | |
82 | ||
83 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
84 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
85 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
86 | ||
87 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
88 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
89 | { | |
90 | if (vect_print_dump_info (REPORT_DETAILS)) | |
91 | fprintf (vect_dump, "already marked relevant/live."); | |
92 | return; | |
93 | } | |
94 | ||
95 | VEC_safe_push (gimple, heap, *worklist, stmt); | |
96 | } | |
97 | ||
98 | ||
99 | /* Function vect_stmt_relevant_p. | |
100 | ||
101 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
102 | "relevant for vectorization". | |
103 | ||
104 | A stmt is considered "relevant for vectorization" if: | |
105 | - it has uses outside the loop. | |
106 | - it has vdefs (it alters memory). | |
107 | - control stmts in the loop (except for the exit condition). | |
108 | ||
109 | CHECKME: what other side effects would the vectorizer allow? */ | |
110 | ||
111 | static bool | |
112 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
113 | enum vect_relevant *relevant, bool *live_p) | |
114 | { | |
115 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
116 | ssa_op_iter op_iter; | |
117 | imm_use_iterator imm_iter; | |
118 | use_operand_p use_p; | |
119 | def_operand_p def_p; | |
120 | ||
8644a673 | 121 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
122 | *live_p = false; |
123 | ||
124 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
125 | if (is_ctrl_stmt (stmt) |
126 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
127 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 128 | *relevant = vect_used_in_scope; |
ebfd146a IR |
129 | |
130 | /* changing memory. */ | |
131 | if (gimple_code (stmt) != GIMPLE_PHI) | |
5006671f | 132 | if (gimple_vdef (stmt)) |
ebfd146a IR |
133 | { |
134 | if (vect_print_dump_info (REPORT_DETAILS)) | |
135 | fprintf (vect_dump, "vec_stmt_relevant_p: stmt has vdefs."); | |
8644a673 | 136 | *relevant = vect_used_in_scope; |
ebfd146a IR |
137 | } |
138 | ||
139 | /* uses outside the loop. */ | |
140 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
141 | { | |
142 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
143 | { | |
144 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
145 | if (!flow_bb_inside_loop_p (loop, bb)) | |
146 | { | |
147 | if (vect_print_dump_info (REPORT_DETAILS)) | |
148 | fprintf (vect_dump, "vec_stmt_relevant_p: used out of loop."); | |
149 | ||
3157b0c2 AO |
150 | if (is_gimple_debug (USE_STMT (use_p))) |
151 | continue; | |
152 | ||
ebfd146a IR |
153 | /* We expect all such uses to be in the loop exit phis |
154 | (because of loop closed form) */ | |
155 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
156 | gcc_assert (bb == single_exit (loop)->dest); | |
157 | ||
158 | *live_p = true; | |
159 | } | |
160 | } | |
161 | } | |
162 | ||
163 | return (*live_p || *relevant); | |
164 | } | |
165 | ||
166 | ||
b8698a0f | 167 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 168 | |
b8698a0f | 169 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
170 | used in STMT for anything other than indexing an array. */ |
171 | ||
172 | static bool | |
173 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
174 | { | |
175 | tree operand; | |
176 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 177 | |
ebfd146a IR |
178 | /* USE corresponds to some operand in STMT. If there is no data |
179 | reference in STMT, then any operand that corresponds to USE | |
180 | is not indexing an array. */ | |
181 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
182 | return true; | |
59a05b0c | 183 | |
ebfd146a IR |
184 | /* STMT has a data_ref. FORNOW this means that its of one of |
185 | the following forms: | |
186 | -1- ARRAY_REF = var | |
187 | -2- var = ARRAY_REF | |
188 | (This should have been verified in analyze_data_refs). | |
189 | ||
190 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 191 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
192 | for array indexing. |
193 | ||
194 | Therefore, all we need to check is if STMT falls into the | |
195 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
196 | |
197 | if (!gimple_assign_copy_p (stmt)) | |
198 | return false; | |
59a05b0c EB |
199 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
200 | return false; | |
ebfd146a | 201 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
202 | if (TREE_CODE (operand) != SSA_NAME) |
203 | return false; | |
204 | ||
205 | if (operand == use) | |
206 | return true; | |
207 | ||
208 | return false; | |
209 | } | |
210 | ||
211 | ||
b8698a0f | 212 | /* |
ebfd146a IR |
213 | Function process_use. |
214 | ||
215 | Inputs: | |
216 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 217 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ebfd146a IR |
218 | that defined USE. This is done by calling mark_relevant and passing it |
219 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). | |
220 | ||
221 | Outputs: | |
222 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
223 | relevance info of the DEF_STMT of this USE: | |
224 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
225 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
226 | Exceptions: | |
227 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 228 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 229 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
230 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
231 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
232 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
233 | be modified accordingly. | |
234 | ||
235 | Return true if everything is as expected. Return false otherwise. */ | |
236 | ||
237 | static bool | |
b8698a0f | 238 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
ebfd146a IR |
239 | enum vect_relevant relevant, VEC(gimple,heap) **worklist) |
240 | { | |
241 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
242 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
243 | stmt_vec_info dstmt_vinfo; | |
244 | basic_block bb, def_bb; | |
245 | tree def; | |
246 | gimple def_stmt; | |
247 | enum vect_def_type dt; | |
248 | ||
b8698a0f | 249 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a IR |
250 | that are used for address computation are not considered relevant. */ |
251 | if (!exist_non_indexing_operands_for_use_p (use, stmt)) | |
252 | return true; | |
253 | ||
a70d6342 | 254 | if (!vect_is_simple_use (use, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
b8698a0f | 255 | { |
8644a673 | 256 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
257 | fprintf (vect_dump, "not vectorized: unsupported use in stmt."); |
258 | return false; | |
259 | } | |
260 | ||
261 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
262 | return true; | |
263 | ||
264 | def_bb = gimple_bb (def_stmt); | |
265 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
266 | { | |
267 | if (vect_print_dump_info (REPORT_DETAILS)) | |
268 | fprintf (vect_dump, "def_stmt is out of loop."); | |
269 | return true; | |
270 | } | |
271 | ||
b8698a0f L |
272 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
273 | DEF_STMT must have already been processed, because this should be the | |
274 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
275 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
276 | check that everything is as expected, and we are done. */ |
277 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
278 | bb = gimple_bb (stmt); | |
279 | if (gimple_code (stmt) == GIMPLE_PHI | |
280 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
281 | && gimple_code (def_stmt) != GIMPLE_PHI | |
282 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
283 | && bb->loop_father == def_bb->loop_father) | |
284 | { | |
285 | if (vect_print_dump_info (REPORT_DETAILS)) | |
286 | fprintf (vect_dump, "reduc-stmt defining reduc-phi in the same nest."); | |
287 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) | |
288 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
289 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 290 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 291 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
292 | return true; |
293 | } | |
294 | ||
295 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
296 | outer-loop-header-bb: | |
297 | d = def_stmt | |
298 | inner-loop: | |
299 | stmt # use (d) | |
300 | outer-loop-tail-bb: | |
301 | ... */ | |
302 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
303 | { | |
304 | if (vect_print_dump_info (REPORT_DETAILS)) | |
305 | fprintf (vect_dump, "outer-loop def-stmt defining inner-loop stmt."); | |
7c5222ff | 306 | |
ebfd146a IR |
307 | switch (relevant) |
308 | { | |
8644a673 | 309 | case vect_unused_in_scope: |
7c5222ff IR |
310 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
311 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 312 | break; |
7c5222ff | 313 | |
ebfd146a | 314 | case vect_used_in_outer_by_reduction: |
7c5222ff | 315 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
316 | relevant = vect_used_by_reduction; |
317 | break; | |
7c5222ff | 318 | |
ebfd146a | 319 | case vect_used_in_outer: |
7c5222ff | 320 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 321 | relevant = vect_used_in_scope; |
ebfd146a | 322 | break; |
7c5222ff | 323 | |
8644a673 | 324 | case vect_used_in_scope: |
ebfd146a IR |
325 | break; |
326 | ||
327 | default: | |
328 | gcc_unreachable (); | |
b8698a0f | 329 | } |
ebfd146a IR |
330 | } |
331 | ||
332 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
333 | outer-loop-header-bb: | |
334 | ... | |
335 | inner-loop: | |
336 | d = def_stmt | |
06066f92 | 337 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
338 | stmt # use (d) */ |
339 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
340 | { | |
341 | if (vect_print_dump_info (REPORT_DETAILS)) | |
342 | fprintf (vect_dump, "inner-loop def-stmt defining outer-loop stmt."); | |
7c5222ff | 343 | |
ebfd146a IR |
344 | switch (relevant) |
345 | { | |
8644a673 | 346 | case vect_unused_in_scope: |
b8698a0f | 347 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 348 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 349 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
350 | break; |
351 | ||
ebfd146a IR |
352 | case vect_used_by_reduction: |
353 | relevant = vect_used_in_outer_by_reduction; | |
354 | break; | |
355 | ||
8644a673 | 356 | case vect_used_in_scope: |
ebfd146a IR |
357 | relevant = vect_used_in_outer; |
358 | break; | |
359 | ||
360 | default: | |
361 | gcc_unreachable (); | |
362 | } | |
363 | } | |
364 | ||
365 | vect_mark_relevant (worklist, def_stmt, relevant, live_p); | |
366 | return true; | |
367 | } | |
368 | ||
369 | ||
370 | /* Function vect_mark_stmts_to_be_vectorized. | |
371 | ||
372 | Not all stmts in the loop need to be vectorized. For example: | |
373 | ||
374 | for i... | |
375 | for j... | |
376 | 1. T0 = i + j | |
377 | 2. T1 = a[T0] | |
378 | ||
379 | 3. j = j + 1 | |
380 | ||
381 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
382 | addressing of vectorized data-refs are handled differently. | |
383 | ||
384 | This pass detects such stmts. */ | |
385 | ||
386 | bool | |
387 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
388 | { | |
389 | VEC(gimple,heap) *worklist; | |
390 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
391 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
392 | unsigned int nbbs = loop->num_nodes; | |
393 | gimple_stmt_iterator si; | |
394 | gimple stmt; | |
395 | unsigned int i; | |
396 | stmt_vec_info stmt_vinfo; | |
397 | basic_block bb; | |
398 | gimple phi; | |
399 | bool live_p; | |
06066f92 IR |
400 | enum vect_relevant relevant, tmp_relevant; |
401 | enum vect_def_type def_type; | |
ebfd146a IR |
402 | |
403 | if (vect_print_dump_info (REPORT_DETAILS)) | |
404 | fprintf (vect_dump, "=== vect_mark_stmts_to_be_vectorized ==="); | |
405 | ||
406 | worklist = VEC_alloc (gimple, heap, 64); | |
407 | ||
408 | /* 1. Init worklist. */ | |
409 | for (i = 0; i < nbbs; i++) | |
410 | { | |
411 | bb = bbs[i]; | |
412 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 413 | { |
ebfd146a IR |
414 | phi = gsi_stmt (si); |
415 | if (vect_print_dump_info (REPORT_DETAILS)) | |
416 | { | |
417 | fprintf (vect_dump, "init: phi relevant? "); | |
418 | print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); | |
419 | } | |
420 | ||
421 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
422 | vect_mark_relevant (&worklist, phi, relevant, live_p); | |
423 | } | |
424 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
425 | { | |
426 | stmt = gsi_stmt (si); | |
427 | if (vect_print_dump_info (REPORT_DETAILS)) | |
428 | { | |
429 | fprintf (vect_dump, "init: stmt relevant? "); | |
430 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
b8698a0f | 431 | } |
ebfd146a IR |
432 | |
433 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
434 | vect_mark_relevant (&worklist, stmt, relevant, live_p); | |
435 | } | |
436 | } | |
437 | ||
438 | /* 2. Process_worklist */ | |
439 | while (VEC_length (gimple, worklist) > 0) | |
440 | { | |
441 | use_operand_p use_p; | |
442 | ssa_op_iter iter; | |
443 | ||
444 | stmt = VEC_pop (gimple, worklist); | |
445 | if (vect_print_dump_info (REPORT_DETAILS)) | |
446 | { | |
447 | fprintf (vect_dump, "worklist: examine stmt: "); | |
448 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
449 | } | |
450 | ||
b8698a0f L |
451 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
452 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
453 | liveness and relevance properties of STMT. */ |
454 | stmt_vinfo = vinfo_for_stmt (stmt); | |
455 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
456 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
457 | ||
458 | /* Generally, the liveness and relevance properties of STMT are | |
459 | propagated as is to the DEF_STMTs of its USEs: | |
460 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
461 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
462 | ||
463 | One exception is when STMT has been identified as defining a reduction | |
464 | variable; in this case we set the liveness/relevance as follows: | |
465 | live_p = false | |
466 | relevant = vect_used_by_reduction | |
467 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f L |
468 | those that are used by a reduction computation, and those that are |
469 | (also) used by a regular computation. This allows us later on to | |
470 | identify stmts that are used solely by a reduction, and therefore the | |
7c5222ff | 471 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 472 | |
06066f92 IR |
473 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
474 | tmp_relevant = relevant; | |
475 | switch (def_type) | |
ebfd146a | 476 | { |
06066f92 IR |
477 | case vect_reduction_def: |
478 | switch (tmp_relevant) | |
479 | { | |
480 | case vect_unused_in_scope: | |
481 | relevant = vect_used_by_reduction; | |
482 | break; | |
483 | ||
484 | case vect_used_by_reduction: | |
485 | if (gimple_code (stmt) == GIMPLE_PHI) | |
486 | break; | |
487 | /* fall through */ | |
488 | ||
489 | default: | |
490 | if (vect_print_dump_info (REPORT_DETAILS)) | |
491 | fprintf (vect_dump, "unsupported use of reduction."); | |
492 | ||
493 | VEC_free (gimple, heap, worklist); | |
494 | return false; | |
495 | } | |
496 | ||
b8698a0f | 497 | live_p = false; |
06066f92 | 498 | break; |
b8698a0f | 499 | |
06066f92 IR |
500 | case vect_nested_cycle: |
501 | if (tmp_relevant != vect_unused_in_scope | |
502 | && tmp_relevant != vect_used_in_outer_by_reduction | |
503 | && tmp_relevant != vect_used_in_outer) | |
504 | { | |
505 | if (vect_print_dump_info (REPORT_DETAILS)) | |
506 | fprintf (vect_dump, "unsupported use of nested cycle."); | |
7c5222ff | 507 | |
06066f92 IR |
508 | VEC_free (gimple, heap, worklist); |
509 | return false; | |
510 | } | |
7c5222ff | 511 | |
b8698a0f L |
512 | live_p = false; |
513 | break; | |
514 | ||
06066f92 IR |
515 | case vect_double_reduction_def: |
516 | if (tmp_relevant != vect_unused_in_scope | |
517 | && tmp_relevant != vect_used_by_reduction) | |
518 | { | |
7c5222ff | 519 | if (vect_print_dump_info (REPORT_DETAILS)) |
06066f92 | 520 | fprintf (vect_dump, "unsupported use of double reduction."); |
7c5222ff IR |
521 | |
522 | VEC_free (gimple, heap, worklist); | |
523 | return false; | |
06066f92 IR |
524 | } |
525 | ||
526 | live_p = false; | |
b8698a0f | 527 | break; |
7c5222ff | 528 | |
06066f92 IR |
529 | default: |
530 | break; | |
7c5222ff | 531 | } |
b8698a0f | 532 | |
ebfd146a IR |
533 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) |
534 | { | |
535 | tree op = USE_FROM_PTR (use_p); | |
536 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, &worklist)) | |
537 | { | |
538 | VEC_free (gimple, heap, worklist); | |
539 | return false; | |
540 | } | |
541 | } | |
542 | } /* while worklist */ | |
543 | ||
544 | VEC_free (gimple, heap, worklist); | |
545 | return true; | |
546 | } | |
547 | ||
548 | ||
549 | int | |
550 | cost_for_stmt (gimple stmt) | |
551 | { | |
552 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
553 | ||
554 | switch (STMT_VINFO_TYPE (stmt_info)) | |
555 | { | |
556 | case load_vec_info_type: | |
557 | return TARG_SCALAR_LOAD_COST; | |
558 | case store_vec_info_type: | |
559 | return TARG_SCALAR_STORE_COST; | |
560 | case op_vec_info_type: | |
561 | case condition_vec_info_type: | |
562 | case assignment_vec_info_type: | |
563 | case reduc_vec_info_type: | |
564 | case induc_vec_info_type: | |
565 | case type_promotion_vec_info_type: | |
566 | case type_demotion_vec_info_type: | |
567 | case type_conversion_vec_info_type: | |
568 | case call_vec_info_type: | |
569 | return TARG_SCALAR_STMT_COST; | |
570 | case undef_vec_info_type: | |
571 | default: | |
572 | gcc_unreachable (); | |
573 | } | |
574 | } | |
575 | ||
b8698a0f | 576 | /* Function vect_model_simple_cost. |
ebfd146a | 577 | |
b8698a0f | 578 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
579 | single op. Right now, this does not account for multiple insns that could |
580 | be generated for the single vector op. We will handle that shortly. */ | |
581 | ||
582 | void | |
b8698a0f | 583 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
ebfd146a IR |
584 | enum vect_def_type *dt, slp_tree slp_node) |
585 | { | |
586 | int i; | |
587 | int inside_cost = 0, outside_cost = 0; | |
588 | ||
589 | /* The SLP costs were already calculated during SLP tree build. */ | |
590 | if (PURE_SLP_STMT (stmt_info)) | |
591 | return; | |
592 | ||
593 | inside_cost = ncopies * TARG_VEC_STMT_COST; | |
594 | ||
595 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
596 | for (i = 0; i < 2; i++) | |
597 | { | |
8644a673 | 598 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
b8698a0f | 599 | outside_cost += TARG_SCALAR_TO_VEC_COST; |
ebfd146a | 600 | } |
b8698a0f | 601 | |
ebfd146a IR |
602 | if (vect_print_dump_info (REPORT_COST)) |
603 | fprintf (vect_dump, "vect_model_simple_cost: inside_cost = %d, " | |
604 | "outside_cost = %d .", inside_cost, outside_cost); | |
605 | ||
606 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
607 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
608 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
609 | } | |
610 | ||
611 | ||
b8698a0f L |
612 | /* Function vect_cost_strided_group_size |
613 | ||
ebfd146a IR |
614 | For strided load or store, return the group_size only if it is the first |
615 | load or store of a group, else return 1. This ensures that group size is | |
616 | only returned once per group. */ | |
617 | ||
618 | static int | |
619 | vect_cost_strided_group_size (stmt_vec_info stmt_info) | |
620 | { | |
621 | gimple first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
622 | ||
623 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
624 | return DR_GROUP_SIZE (stmt_info); | |
625 | ||
626 | return 1; | |
627 | } | |
628 | ||
629 | ||
630 | /* Function vect_model_store_cost | |
631 | ||
632 | Models cost for stores. In the case of strided accesses, one access | |
633 | has the overhead of the strided access attributed to it. */ | |
634 | ||
635 | void | |
b8698a0f | 636 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
ebfd146a IR |
637 | enum vect_def_type dt, slp_tree slp_node) |
638 | { | |
639 | int group_size; | |
640 | int inside_cost = 0, outside_cost = 0; | |
641 | ||
642 | /* The SLP costs were already calculated during SLP tree build. */ | |
643 | if (PURE_SLP_STMT (stmt_info)) | |
644 | return; | |
645 | ||
8644a673 | 646 | if (dt == vect_constant_def || dt == vect_external_def) |
ebfd146a IR |
647 | outside_cost = TARG_SCALAR_TO_VEC_COST; |
648 | ||
649 | /* Strided access? */ | |
b8698a0f | 650 | if (DR_GROUP_FIRST_DR (stmt_info) && !slp_node) |
ebfd146a IR |
651 | group_size = vect_cost_strided_group_size (stmt_info); |
652 | /* Not a strided access. */ | |
653 | else | |
654 | group_size = 1; | |
655 | ||
b8698a0f | 656 | /* Is this an access in a group of stores, which provide strided access? |
ebfd146a | 657 | If so, add in the cost of the permutes. */ |
b8698a0f | 658 | if (group_size > 1) |
ebfd146a IR |
659 | { |
660 | /* Uses a high and low interleave operation for each needed permute. */ | |
b8698a0f | 661 | inside_cost = ncopies * exact_log2(group_size) * group_size |
ebfd146a IR |
662 | * TARG_VEC_STMT_COST; |
663 | ||
664 | if (vect_print_dump_info (REPORT_COST)) | |
665 | fprintf (vect_dump, "vect_model_store_cost: strided group_size = %d .", | |
666 | group_size); | |
667 | ||
668 | } | |
669 | ||
670 | /* Costs of the stores. */ | |
671 | inside_cost += ncopies * TARG_VEC_STORE_COST; | |
672 | ||
673 | if (vect_print_dump_info (REPORT_COST)) | |
674 | fprintf (vect_dump, "vect_model_store_cost: inside_cost = %d, " | |
675 | "outside_cost = %d .", inside_cost, outside_cost); | |
676 | ||
677 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
678 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
679 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
680 | } | |
681 | ||
682 | ||
683 | /* Function vect_model_load_cost | |
684 | ||
685 | Models cost for loads. In the case of strided accesses, the last access | |
686 | has the overhead of the strided access attributed to it. Since unaligned | |
b8698a0f | 687 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
688 | access scheme chosen. */ |
689 | ||
690 | void | |
691 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, slp_tree slp_node) | |
b8698a0f | 692 | |
ebfd146a IR |
693 | { |
694 | int group_size; | |
695 | int alignment_support_cheme; | |
696 | gimple first_stmt; | |
697 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
698 | int inside_cost = 0, outside_cost = 0; | |
699 | ||
700 | /* The SLP costs were already calculated during SLP tree build. */ | |
701 | if (PURE_SLP_STMT (stmt_info)) | |
702 | return; | |
703 | ||
704 | /* Strided accesses? */ | |
705 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
706 | if (first_stmt && !slp_node) | |
707 | { | |
708 | group_size = vect_cost_strided_group_size (stmt_info); | |
709 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
710 | } | |
711 | /* Not a strided access. */ | |
712 | else | |
713 | { | |
714 | group_size = 1; | |
715 | first_dr = dr; | |
716 | } | |
717 | ||
718 | alignment_support_cheme = vect_supportable_dr_alignment (first_dr); | |
719 | ||
b8698a0f | 720 | /* Is this an access in a group of loads providing strided access? |
ebfd146a | 721 | If so, add in the cost of the permutes. */ |
b8698a0f | 722 | if (group_size > 1) |
ebfd146a IR |
723 | { |
724 | /* Uses an even and odd extract operations for each needed permute. */ | |
725 | inside_cost = ncopies * exact_log2(group_size) * group_size | |
726 | * TARG_VEC_STMT_COST; | |
727 | ||
728 | if (vect_print_dump_info (REPORT_COST)) | |
729 | fprintf (vect_dump, "vect_model_load_cost: strided group_size = %d .", | |
730 | group_size); | |
731 | ||
732 | } | |
733 | ||
734 | /* The loads themselves. */ | |
735 | switch (alignment_support_cheme) | |
736 | { | |
737 | case dr_aligned: | |
738 | { | |
739 | inside_cost += ncopies * TARG_VEC_LOAD_COST; | |
740 | ||
741 | if (vect_print_dump_info (REPORT_COST)) | |
742 | fprintf (vect_dump, "vect_model_load_cost: aligned."); | |
743 | ||
744 | break; | |
745 | } | |
746 | case dr_unaligned_supported: | |
747 | { | |
748 | /* Here, we assign an additional cost for the unaligned load. */ | |
749 | inside_cost += ncopies * TARG_VEC_UNALIGNED_LOAD_COST; | |
750 | ||
751 | if (vect_print_dump_info (REPORT_COST)) | |
752 | fprintf (vect_dump, "vect_model_load_cost: unaligned supported by " | |
753 | "hardware."); | |
754 | ||
755 | break; | |
756 | } | |
757 | case dr_explicit_realign: | |
758 | { | |
759 | inside_cost += ncopies * (2*TARG_VEC_LOAD_COST + TARG_VEC_STMT_COST); | |
760 | ||
761 | /* FIXME: If the misalignment remains fixed across the iterations of | |
762 | the containing loop, the following cost should be added to the | |
763 | outside costs. */ | |
764 | if (targetm.vectorize.builtin_mask_for_load) | |
765 | inside_cost += TARG_VEC_STMT_COST; | |
766 | ||
767 | break; | |
768 | } | |
769 | case dr_explicit_realign_optimized: | |
770 | { | |
771 | if (vect_print_dump_info (REPORT_COST)) | |
772 | fprintf (vect_dump, "vect_model_load_cost: unaligned software " | |
773 | "pipelined."); | |
774 | ||
775 | /* Unaligned software pipeline has a load of an address, an initial | |
776 | load, and possibly a mask operation to "prime" the loop. However, | |
777 | if this is an access in a group of loads, which provide strided | |
778 | access, then the above cost should only be considered for one | |
779 | access in the group. Inside the loop, there is a load op | |
780 | and a realignment op. */ | |
781 | ||
782 | if ((!DR_GROUP_FIRST_DR (stmt_info)) || group_size > 1 || slp_node) | |
783 | { | |
784 | outside_cost = 2*TARG_VEC_STMT_COST; | |
785 | if (targetm.vectorize.builtin_mask_for_load) | |
786 | outside_cost += TARG_VEC_STMT_COST; | |
787 | } | |
788 | ||
789 | inside_cost += ncopies * (TARG_VEC_LOAD_COST + TARG_VEC_STMT_COST); | |
790 | ||
791 | break; | |
792 | } | |
793 | ||
794 | default: | |
795 | gcc_unreachable (); | |
796 | } | |
b8698a0f | 797 | |
ebfd146a IR |
798 | if (vect_print_dump_info (REPORT_COST)) |
799 | fprintf (vect_dump, "vect_model_load_cost: inside_cost = %d, " | |
800 | "outside_cost = %d .", inside_cost, outside_cost); | |
801 | ||
802 | /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */ | |
803 | stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost); | |
804 | stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost); | |
805 | } | |
806 | ||
807 | ||
808 | /* Function vect_init_vector. | |
809 | ||
810 | Insert a new stmt (INIT_STMT) that initializes a new vector variable with | |
811 | the vector elements of VECTOR_VAR. Place the initialization at BSI if it | |
812 | is not NULL. Otherwise, place the initialization at the loop preheader. | |
b8698a0f | 813 | Return the DEF of INIT_STMT. |
ebfd146a IR |
814 | It will be used in the vectorization of STMT. */ |
815 | ||
816 | tree | |
817 | vect_init_vector (gimple stmt, tree vector_var, tree vector_type, | |
818 | gimple_stmt_iterator *gsi) | |
819 | { | |
820 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
821 | tree new_var; | |
822 | gimple init_stmt; | |
823 | tree vec_oprnd; | |
824 | edge pe; | |
825 | tree new_temp; | |
826 | basic_block new_bb; | |
b8698a0f | 827 | |
ebfd146a | 828 | new_var = vect_get_new_vect_var (vector_type, vect_simple_var, "cst_"); |
b8698a0f | 829 | add_referenced_var (new_var); |
ebfd146a IR |
830 | init_stmt = gimple_build_assign (new_var, vector_var); |
831 | new_temp = make_ssa_name (new_var, init_stmt); | |
832 | gimple_assign_set_lhs (init_stmt, new_temp); | |
833 | ||
834 | if (gsi) | |
835 | vect_finish_stmt_generation (stmt, init_stmt, gsi); | |
836 | else | |
837 | { | |
838 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
b8698a0f | 839 | |
a70d6342 IR |
840 | if (loop_vinfo) |
841 | { | |
842 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
843 | ||
844 | if (nested_in_vect_loop_p (loop, stmt)) | |
845 | loop = loop->inner; | |
b8698a0f | 846 | |
a70d6342 IR |
847 | pe = loop_preheader_edge (loop); |
848 | new_bb = gsi_insert_on_edge_immediate (pe, init_stmt); | |
849 | gcc_assert (!new_bb); | |
850 | } | |
851 | else | |
852 | { | |
853 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
854 | basic_block bb; | |
855 | gimple_stmt_iterator gsi_bb_start; | |
856 | ||
857 | gcc_assert (bb_vinfo); | |
858 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 859 | gsi_bb_start = gsi_after_labels (bb); |
a70d6342 IR |
860 | gsi_insert_before (&gsi_bb_start, init_stmt, GSI_SAME_STMT); |
861 | } | |
ebfd146a IR |
862 | } |
863 | ||
864 | if (vect_print_dump_info (REPORT_DETAILS)) | |
865 | { | |
866 | fprintf (vect_dump, "created new init_stmt: "); | |
867 | print_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM); | |
868 | } | |
869 | ||
870 | vec_oprnd = gimple_assign_lhs (init_stmt); | |
871 | return vec_oprnd; | |
872 | } | |
873 | ||
a70d6342 | 874 | |
ebfd146a IR |
875 | /* Function vect_get_vec_def_for_operand. |
876 | ||
877 | OP is an operand in STMT. This function returns a (vector) def that will be | |
878 | used in the vectorized stmt for STMT. | |
879 | ||
880 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
881 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
882 | ||
883 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
884 | needs to be introduced. */ | |
885 | ||
886 | tree | |
887 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
888 | { | |
889 | tree vec_oprnd; | |
890 | gimple vec_stmt; | |
891 | gimple def_stmt; | |
892 | stmt_vec_info def_stmt_info = NULL; | |
893 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
894 | tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
895 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
896 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
897 | tree vec_inv; | |
898 | tree vec_cst; | |
899 | tree t = NULL_TREE; | |
900 | tree def; | |
901 | int i; | |
902 | enum vect_def_type dt; | |
903 | bool is_simple_use; | |
904 | tree vector_type; | |
905 | ||
906 | if (vect_print_dump_info (REPORT_DETAILS)) | |
907 | { | |
908 | fprintf (vect_dump, "vect_get_vec_def_for_operand: "); | |
909 | print_generic_expr (vect_dump, op, TDF_SLIM); | |
910 | } | |
911 | ||
b8698a0f | 912 | is_simple_use = vect_is_simple_use (op, loop_vinfo, NULL, &def_stmt, &def, |
a70d6342 | 913 | &dt); |
ebfd146a IR |
914 | gcc_assert (is_simple_use); |
915 | if (vect_print_dump_info (REPORT_DETAILS)) | |
916 | { | |
917 | if (def) | |
918 | { | |
919 | fprintf (vect_dump, "def = "); | |
920 | print_generic_expr (vect_dump, def, TDF_SLIM); | |
921 | } | |
922 | if (def_stmt) | |
923 | { | |
924 | fprintf (vect_dump, " def_stmt = "); | |
925 | print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM); | |
926 | } | |
927 | } | |
928 | ||
929 | switch (dt) | |
930 | { | |
931 | /* Case 1: operand is a constant. */ | |
932 | case vect_constant_def: | |
933 | { | |
7569a6cc RG |
934 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
935 | gcc_assert (vector_type); | |
936 | ||
b8698a0f | 937 | if (scalar_def) |
ebfd146a IR |
938 | *scalar_def = op; |
939 | ||
940 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
941 | if (vect_print_dump_info (REPORT_DETAILS)) | |
942 | fprintf (vect_dump, "Create vector_cst. nunits = %d", nunits); | |
943 | ||
944 | for (i = nunits - 1; i >= 0; --i) | |
945 | { | |
946 | t = tree_cons (NULL_TREE, op, t); | |
947 | } | |
7569a6cc RG |
948 | vec_cst = build_vector (vector_type, t); |
949 | return vect_init_vector (stmt, vec_cst, vector_type, NULL); | |
ebfd146a IR |
950 | } |
951 | ||
952 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 953 | case vect_external_def: |
ebfd146a IR |
954 | { |
955 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
956 | gcc_assert (vector_type); | |
957 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); | |
958 | ||
b8698a0f | 959 | if (scalar_def) |
ebfd146a IR |
960 | *scalar_def = def; |
961 | ||
962 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
963 | if (vect_print_dump_info (REPORT_DETAILS)) | |
964 | fprintf (vect_dump, "Create vector_inv."); | |
965 | ||
966 | for (i = nunits - 1; i >= 0; --i) | |
967 | { | |
968 | t = tree_cons (NULL_TREE, def, t); | |
969 | } | |
970 | ||
971 | /* FIXME: use build_constructor directly. */ | |
972 | vec_inv = build_constructor_from_list (vector_type, t); | |
973 | return vect_init_vector (stmt, vec_inv, vector_type, NULL); | |
974 | } | |
975 | ||
976 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 977 | case vect_internal_def: |
ebfd146a | 978 | { |
b8698a0f | 979 | if (scalar_def) |
ebfd146a IR |
980 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; |
981 | ||
982 | /* Get the def from the vectorized stmt. */ | |
983 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
984 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
985 | gcc_assert (vec_stmt); | |
986 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
987 | vec_oprnd = PHI_RESULT (vec_stmt); | |
988 | else if (is_gimple_call (vec_stmt)) | |
989 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
990 | else | |
991 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
992 | return vec_oprnd; | |
993 | } | |
994 | ||
995 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
996 | case vect_reduction_def: | |
06066f92 | 997 | case vect_double_reduction_def: |
7c5222ff | 998 | case vect_nested_cycle: |
ebfd146a IR |
999 | { |
1000 | struct loop *loop; | |
1001 | ||
1002 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
b8698a0f | 1003 | loop = (gimple_bb (def_stmt))->loop_father; |
ebfd146a IR |
1004 | |
1005 | /* Get the def before the loop */ | |
1006 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1007 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1008 | } | |
1009 | ||
1010 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1011 | case vect_induction_def: | |
1012 | { | |
1013 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1014 | ||
1015 | /* Get the def from the vectorized stmt. */ | |
1016 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1017 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
1018 | gcc_assert (vec_stmt && gimple_code (vec_stmt) == GIMPLE_PHI); | |
1019 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1020 | return vec_oprnd; | |
1021 | } | |
1022 | ||
1023 | default: | |
1024 | gcc_unreachable (); | |
1025 | } | |
1026 | } | |
1027 | ||
1028 | ||
1029 | /* Function vect_get_vec_def_for_stmt_copy | |
1030 | ||
b8698a0f L |
1031 | Return a vector-def for an operand. This function is used when the |
1032 | vectorized stmt to be created (by the caller to this function) is a "copy" | |
1033 | created in case the vectorized result cannot fit in one vector, and several | |
1034 | copies of the vector-stmt are required. In this case the vector-def is | |
ebfd146a | 1035 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1036 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1037 | DT is the type of the vector def VEC_OPRND. |
1038 | ||
1039 | Context: | |
1040 | In case the vectorization factor (VF) is bigger than the number | |
1041 | of elements that can fit in a vectype (nunits), we have to generate | |
1042 | more than one vector stmt to vectorize the scalar stmt. This situation | |
b8698a0f | 1043 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1044 | smallest data-type determines the VF, and as a result, when vectorizing |
1045 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1046 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1047 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1048 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1049 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1050 | ||
1051 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1052 | |
ebfd146a IR |
1053 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1054 | VS1.1: vx.1 = memref1 VS1.2 | |
1055 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1056 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1057 | |
1058 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1059 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1060 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1061 | VSnew.3: vz3 = vx.3 + ... | |
1062 | ||
1063 | The vectorization of S1 is explained in vectorizable_load. | |
1064 | The vectorization of S2: | |
b8698a0f L |
1065 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1066 | the function 'vect_get_vec_def_for_operand' is called to | |
ebfd146a IR |
1067 | get the relevant vector-def for each operand of S2. For operand x it |
1068 | returns the vector-def 'vx.0'. | |
1069 | ||
b8698a0f L |
1070 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1071 | function is called to get the relevant vector-def for each operand. It is | |
1072 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1073 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1074 | ||
b8698a0f L |
1075 | For example, to obtain the vector-def 'vx.1' in order to create the |
1076 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1077 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1078 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1079 | and return its def ('vx.1'). | |
1080 | Overall, to create the above sequence this function will be called 3 times: | |
1081 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1082 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1083 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1084 | ||
1085 | tree | |
1086 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1087 | { | |
1088 | gimple vec_stmt_for_operand; | |
1089 | stmt_vec_info def_stmt_info; | |
1090 | ||
1091 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1092 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1093 | return vec_oprnd; |
1094 | ||
1095 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1096 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1097 | gcc_assert (def_stmt_info); | |
1098 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1099 | gcc_assert (vec_stmt_for_operand); | |
1100 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1101 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1102 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1103 | else | |
1104 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1105 | return vec_oprnd; | |
1106 | } | |
1107 | ||
1108 | ||
1109 | /* Get vectorized definitions for the operands to create a copy of an original | |
1110 | stmt. See vect_get_vec_def_for_stmt_copy() for details. */ | |
1111 | ||
1112 | static void | |
b8698a0f L |
1113 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
1114 | VEC(tree,heap) **vec_oprnds0, | |
ebfd146a IR |
1115 | VEC(tree,heap) **vec_oprnds1) |
1116 | { | |
1117 | tree vec_oprnd = VEC_pop (tree, *vec_oprnds0); | |
1118 | ||
1119 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
1120 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); | |
1121 | ||
1122 | if (vec_oprnds1 && *vec_oprnds1) | |
1123 | { | |
1124 | vec_oprnd = VEC_pop (tree, *vec_oprnds1); | |
1125 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); | |
1126 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); | |
1127 | } | |
1128 | } | |
1129 | ||
1130 | ||
1131 | /* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not NULL. */ | |
1132 | ||
1133 | static void | |
1134 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, | |
1135 | VEC(tree,heap) **vec_oprnds0, VEC(tree,heap) **vec_oprnds1, | |
1136 | slp_tree slp_node) | |
1137 | { | |
1138 | if (slp_node) | |
b5aeb3bb | 1139 | vect_get_slp_defs (slp_node, vec_oprnds0, vec_oprnds1, -1); |
ebfd146a IR |
1140 | else |
1141 | { | |
1142 | tree vec_oprnd; | |
1143 | ||
b8698a0f L |
1144 | *vec_oprnds0 = VEC_alloc (tree, heap, 1); |
1145 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
ebfd146a IR |
1146 | VEC_quick_push (tree, *vec_oprnds0, vec_oprnd); |
1147 | ||
1148 | if (op1) | |
1149 | { | |
b8698a0f L |
1150 | *vec_oprnds1 = VEC_alloc (tree, heap, 1); |
1151 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
ebfd146a IR |
1152 | VEC_quick_push (tree, *vec_oprnds1, vec_oprnd); |
1153 | } | |
1154 | } | |
1155 | } | |
1156 | ||
1157 | ||
1158 | /* Function vect_finish_stmt_generation. | |
1159 | ||
1160 | Insert a new stmt. */ | |
1161 | ||
1162 | void | |
1163 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1164 | gimple_stmt_iterator *gsi) | |
1165 | { | |
1166 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1167 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1168 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1169 | |
1170 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1171 | ||
1172 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); | |
1173 | ||
b8698a0f | 1174 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
a70d6342 | 1175 | bb_vinfo)); |
ebfd146a IR |
1176 | |
1177 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1178 | { | |
1179 | fprintf (vect_dump, "add new stmt: "); | |
1180 | print_gimple_stmt (vect_dump, vec_stmt, 0, TDF_SLIM); | |
1181 | } | |
1182 | ||
1183 | gimple_set_location (vec_stmt, gimple_location (gsi_stmt (*gsi))); | |
1184 | } | |
1185 | ||
1186 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1187 | a function declaration if the target has a vectorized version | |
1188 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1189 | ||
1190 | tree | |
1191 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1192 | { | |
1193 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1194 | |
1195 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1196 | const or novops ones. */ | |
1197 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1198 | return NULL_TREE; | |
1199 | ||
1200 | if (!fndecl | |
1201 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1202 | || !DECL_BUILT_IN (fndecl)) | |
1203 | return NULL_TREE; | |
1204 | ||
62f7fd21 | 1205 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1206 | vectype_in); |
1207 | } | |
1208 | ||
1209 | /* Function vectorizable_call. | |
1210 | ||
b8698a0f L |
1211 | Check if STMT performs a function call that can be vectorized. |
1212 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1213 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1214 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1215 | ||
1216 | static bool | |
1217 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt) | |
1218 | { | |
1219 | tree vec_dest; | |
1220 | tree scalar_dest; | |
1221 | tree op, type; | |
1222 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1223 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
1224 | tree vectype_out, vectype_in; | |
1225 | int nunits_in; | |
1226 | int nunits_out; | |
1227 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
b690cc0f | 1228 | tree fndecl, new_temp, def, rhs_type; |
ebfd146a IR |
1229 | gimple def_stmt; |
1230 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 1231 | gimple new_stmt = NULL; |
ebfd146a IR |
1232 | int ncopies, j; |
1233 | VEC(tree, heap) *vargs = NULL; | |
1234 | enum { NARROW, NONE, WIDEN } modifier; | |
1235 | size_t i, nargs; | |
1236 | ||
a70d6342 IR |
1237 | /* FORNOW: unsupported in basic block SLP. */ |
1238 | gcc_assert (loop_vinfo); | |
b8698a0f | 1239 | |
ebfd146a IR |
1240 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
1241 | return false; | |
1242 | ||
8644a673 | 1243 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1244 | return false; |
1245 | ||
1246 | /* FORNOW: SLP not supported. */ | |
1247 | if (STMT_SLP_TYPE (stmt_info)) | |
1248 | return false; | |
1249 | ||
1250 | /* Is STMT a vectorizable call? */ | |
1251 | if (!is_gimple_call (stmt)) | |
1252 | return false; | |
1253 | ||
1254 | if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
1255 | return false; | |
1256 | ||
b690cc0f RG |
1257 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
1258 | ||
ebfd146a IR |
1259 | /* Process function arguments. */ |
1260 | rhs_type = NULL_TREE; | |
b690cc0f | 1261 | vectype_in = NULL_TREE; |
ebfd146a IR |
1262 | nargs = gimple_call_num_args (stmt); |
1263 | ||
1264 | /* Bail out if the function has more than two arguments, we | |
1265 | do not have interesting builtin functions to vectorize with | |
1266 | more than two arguments. No arguments is also not good. */ | |
1267 | if (nargs == 0 || nargs > 2) | |
1268 | return false; | |
1269 | ||
1270 | for (i = 0; i < nargs; i++) | |
1271 | { | |
b690cc0f RG |
1272 | tree opvectype; |
1273 | ||
ebfd146a IR |
1274 | op = gimple_call_arg (stmt, i); |
1275 | ||
1276 | /* We can only handle calls with arguments of the same type. */ | |
1277 | if (rhs_type | |
8533c9d8 | 1278 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a IR |
1279 | { |
1280 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1281 | fprintf (vect_dump, "argument types differ."); | |
1282 | return false; | |
1283 | } | |
b690cc0f RG |
1284 | if (!rhs_type) |
1285 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 1286 | |
b690cc0f RG |
1287 | if (!vect_is_simple_use_1 (op, loop_vinfo, NULL, |
1288 | &def_stmt, &def, &dt[i], &opvectype)) | |
ebfd146a IR |
1289 | { |
1290 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1291 | fprintf (vect_dump, "use not simple."); | |
1292 | return false; | |
1293 | } | |
ebfd146a | 1294 | |
b690cc0f RG |
1295 | if (!vectype_in) |
1296 | vectype_in = opvectype; | |
1297 | else if (opvectype | |
1298 | && opvectype != vectype_in) | |
1299 | { | |
1300 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1301 | fprintf (vect_dump, "argument vector types differ."); | |
1302 | return false; | |
1303 | } | |
1304 | } | |
1305 | /* If all arguments are external or constant defs use a vector type with | |
1306 | the same size as the output vector type. */ | |
ebfd146a | 1307 | if (!vectype_in) |
b690cc0f | 1308 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1309 | if (vec_stmt) |
1310 | gcc_assert (vectype_in); | |
1311 | if (!vectype_in) | |
1312 | { | |
1313 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1314 | { | |
1315 | fprintf (vect_dump, "no vectype for scalar type "); | |
1316 | print_generic_expr (vect_dump, rhs_type, TDF_SLIM); | |
1317 | } | |
1318 | ||
1319 | return false; | |
1320 | } | |
ebfd146a IR |
1321 | |
1322 | /* FORNOW */ | |
b690cc0f RG |
1323 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1324 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1325 | if (nunits_in == nunits_out / 2) |
1326 | modifier = NARROW; | |
1327 | else if (nunits_out == nunits_in) | |
1328 | modifier = NONE; | |
1329 | else if (nunits_out == nunits_in / 2) | |
1330 | modifier = WIDEN; | |
1331 | else | |
1332 | return false; | |
1333 | ||
1334 | /* For now, we only vectorize functions if a target specific builtin | |
1335 | is available. TODO -- in some cases, it might be profitable to | |
1336 | insert the calls for pieces of the vector, in order to be able | |
1337 | to vectorize other operations in the loop. */ | |
1338 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
1339 | if (fndecl == NULL_TREE) | |
1340 | { | |
1341 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1342 | fprintf (vect_dump, "function is not vectorizable."); | |
1343 | ||
1344 | return false; | |
1345 | } | |
1346 | ||
5006671f | 1347 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a IR |
1348 | |
1349 | if (modifier == NARROW) | |
1350 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
1351 | else | |
1352 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1353 | ||
1354 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1355 | needs to be generated. */ | |
1356 | gcc_assert (ncopies >= 1); | |
1357 | ||
1358 | if (!vec_stmt) /* transformation not required. */ | |
1359 | { | |
1360 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1361 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1362 | fprintf (vect_dump, "=== vectorizable_call ==="); | |
1363 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
1364 | return true; | |
1365 | } | |
1366 | ||
1367 | /** Transform. **/ | |
1368 | ||
1369 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1370 | fprintf (vect_dump, "transform operation."); | |
1371 | ||
1372 | /* Handle def. */ | |
1373 | scalar_dest = gimple_call_lhs (stmt); | |
1374 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1375 | ||
1376 | prev_stmt_info = NULL; | |
1377 | switch (modifier) | |
1378 | { | |
1379 | case NONE: | |
1380 | for (j = 0; j < ncopies; ++j) | |
1381 | { | |
1382 | /* Build argument list for the vectorized call. */ | |
1383 | if (j == 0) | |
1384 | vargs = VEC_alloc (tree, heap, nargs); | |
1385 | else | |
1386 | VEC_truncate (tree, vargs, 0); | |
1387 | ||
1388 | for (i = 0; i < nargs; i++) | |
1389 | { | |
1390 | op = gimple_call_arg (stmt, i); | |
1391 | if (j == 0) | |
1392 | vec_oprnd0 | |
1393 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1394 | else | |
63827fb8 IR |
1395 | { |
1396 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
1397 | vec_oprnd0 | |
1398 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
1399 | } | |
ebfd146a IR |
1400 | |
1401 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1402 | } | |
1403 | ||
1404 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1405 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1406 | gimple_call_set_lhs (new_stmt, new_temp); | |
1407 | ||
1408 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7411b8f0 | 1409 | mark_symbols_for_renaming (new_stmt); |
ebfd146a IR |
1410 | |
1411 | if (j == 0) | |
1412 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1413 | else | |
1414 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1415 | ||
1416 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1417 | } | |
1418 | ||
1419 | break; | |
1420 | ||
1421 | case NARROW: | |
1422 | for (j = 0; j < ncopies; ++j) | |
1423 | { | |
1424 | /* Build argument list for the vectorized call. */ | |
1425 | if (j == 0) | |
1426 | vargs = VEC_alloc (tree, heap, nargs * 2); | |
1427 | else | |
1428 | VEC_truncate (tree, vargs, 0); | |
1429 | ||
1430 | for (i = 0; i < nargs; i++) | |
1431 | { | |
1432 | op = gimple_call_arg (stmt, i); | |
1433 | if (j == 0) | |
1434 | { | |
1435 | vec_oprnd0 | |
1436 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1437 | vec_oprnd1 | |
63827fb8 | 1438 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1439 | } |
1440 | else | |
1441 | { | |
63827fb8 | 1442 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i); |
ebfd146a | 1443 | vec_oprnd0 |
63827fb8 | 1444 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 1445 | vec_oprnd1 |
63827fb8 | 1446 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
1447 | } |
1448 | ||
1449 | VEC_quick_push (tree, vargs, vec_oprnd0); | |
1450 | VEC_quick_push (tree, vargs, vec_oprnd1); | |
1451 | } | |
1452 | ||
1453 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1454 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1455 | gimple_call_set_lhs (new_stmt, new_temp); | |
1456 | ||
1457 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7411b8f0 | 1458 | mark_symbols_for_renaming (new_stmt); |
ebfd146a IR |
1459 | |
1460 | if (j == 0) | |
1461 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1462 | else | |
1463 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1464 | ||
1465 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1466 | } | |
1467 | ||
1468 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
1469 | ||
1470 | break; | |
1471 | ||
1472 | case WIDEN: | |
1473 | /* No current target implements this case. */ | |
1474 | return false; | |
1475 | } | |
1476 | ||
1477 | VEC_free (tree, heap, vargs); | |
1478 | ||
1479 | /* Update the exception handling table with the vector stmt if necessary. */ | |
1480 | if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt)) | |
1481 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
1482 | ||
1483 | /* The call in STMT might prevent it from being removed in dce. | |
1484 | We however cannot remove it here, due to the way the ssa name | |
1485 | it defines is mapped to the new definition. So just replace | |
1486 | rhs of the statement with something harmless. */ | |
1487 | ||
1488 | type = TREE_TYPE (scalar_dest); | |
1489 | new_stmt = gimple_build_assign (gimple_call_lhs (stmt), | |
1490 | fold_convert (type, integer_zero_node)); | |
1491 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
1492 | set_vinfo_for_stmt (stmt, NULL); | |
1493 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
1494 | gsi_replace (gsi, new_stmt, false); | |
1495 | SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt; | |
1496 | ||
1497 | return true; | |
1498 | } | |
1499 | ||
1500 | ||
1501 | /* Function vect_gen_widened_results_half | |
1502 | ||
1503 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 1504 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ebfd146a IR |
1505 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
1506 | In the case that CODE is a CALL_EXPR, this means that a call to DECL | |
1507 | needs to be created (DECL is a function-decl of a target-builtin). | |
1508 | STMT is the original scalar stmt that we are vectorizing. */ | |
1509 | ||
1510 | static gimple | |
1511 | vect_gen_widened_results_half (enum tree_code code, | |
1512 | tree decl, | |
1513 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
1514 | tree vec_dest, gimple_stmt_iterator *gsi, | |
1515 | gimple stmt) | |
b8698a0f | 1516 | { |
ebfd146a | 1517 | gimple new_stmt; |
b8698a0f L |
1518 | tree new_temp; |
1519 | ||
1520 | /* Generate half of the widened result: */ | |
1521 | if (code == CALL_EXPR) | |
1522 | { | |
1523 | /* Target specific support */ | |
ebfd146a IR |
1524 | if (op_type == binary_op) |
1525 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
1526 | else | |
1527 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
1528 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1529 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
1530 | } |
1531 | else | |
ebfd146a | 1532 | { |
b8698a0f L |
1533 | /* Generic support */ |
1534 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
1535 | if (op_type != binary_op) |
1536 | vec_oprnd1 = NULL; | |
1537 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
1538 | vec_oprnd1); | |
1539 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1540 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 1541 | } |
ebfd146a IR |
1542 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1543 | ||
ebfd146a IR |
1544 | return new_stmt; |
1545 | } | |
1546 | ||
1547 | ||
b8698a0f L |
1548 | /* Check if STMT performs a conversion operation, that can be vectorized. |
1549 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1550 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1551 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1552 | ||
1553 | static bool | |
1554 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
1555 | gimple *vec_stmt, slp_tree slp_node) | |
1556 | { | |
1557 | tree vec_dest; | |
1558 | tree scalar_dest; | |
1559 | tree op0; | |
1560 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1561 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1562 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1563 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
1564 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; | |
1565 | tree new_temp; | |
1566 | tree def; | |
1567 | gimple def_stmt; | |
1568 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
1569 | gimple new_stmt = NULL; | |
1570 | stmt_vec_info prev_stmt_info; | |
1571 | int nunits_in; | |
1572 | int nunits_out; | |
1573 | tree vectype_out, vectype_in; | |
1574 | int ncopies, j; | |
b690cc0f | 1575 | tree rhs_type; |
ebfd146a IR |
1576 | tree builtin_decl; |
1577 | enum { NARROW, NONE, WIDEN } modifier; | |
1578 | int i; | |
1579 | VEC(tree,heap) *vec_oprnds0 = NULL; | |
1580 | tree vop0; | |
ebfd146a IR |
1581 | VEC(tree,heap) *dummy = NULL; |
1582 | int dummy_int; | |
1583 | ||
1584 | /* Is STMT a vectorizable conversion? */ | |
1585 | ||
a70d6342 IR |
1586 | /* FORNOW: unsupported in basic block SLP. */ |
1587 | gcc_assert (loop_vinfo); | |
b8698a0f | 1588 | |
ebfd146a IR |
1589 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
1590 | return false; | |
1591 | ||
8644a673 | 1592 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1593 | return false; |
1594 | ||
1595 | if (!is_gimple_assign (stmt)) | |
1596 | return false; | |
1597 | ||
1598 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
1599 | return false; | |
1600 | ||
1601 | code = gimple_assign_rhs_code (stmt); | |
1602 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
1603 | return false; | |
1604 | ||
1605 | /* Check types of lhs and rhs. */ | |
b690cc0f RG |
1606 | scalar_dest = gimple_assign_lhs (stmt); |
1607 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
1608 | ||
ebfd146a IR |
1609 | op0 = gimple_assign_rhs1 (stmt); |
1610 | rhs_type = TREE_TYPE (op0); | |
b690cc0f RG |
1611 | /* Check the operands of the operation. */ |
1612 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
1613 | &def_stmt, &def, &dt[0], &vectype_in)) | |
1614 | { | |
1615 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1616 | fprintf (vect_dump, "use not simple."); | |
1617 | return false; | |
1618 | } | |
1619 | /* If op0 is an external or constant defs use a vector type of | |
1620 | the same size as the output vector type. */ | |
ebfd146a | 1621 | if (!vectype_in) |
b690cc0f | 1622 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1623 | if (vec_stmt) |
1624 | gcc_assert (vectype_in); | |
1625 | if (!vectype_in) | |
1626 | { | |
1627 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1628 | { | |
1629 | fprintf (vect_dump, "no vectype for scalar type "); | |
1630 | print_generic_expr (vect_dump, rhs_type, TDF_SLIM); | |
1631 | } | |
1632 | ||
1633 | return false; | |
1634 | } | |
ebfd146a IR |
1635 | |
1636 | /* FORNOW */ | |
b690cc0f RG |
1637 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1638 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1639 | if (nunits_in == nunits_out / 2) |
1640 | modifier = NARROW; | |
1641 | else if (nunits_out == nunits_in) | |
1642 | modifier = NONE; | |
1643 | else if (nunits_out == nunits_in / 2) | |
1644 | modifier = WIDEN; | |
1645 | else | |
1646 | return false; | |
1647 | ||
ebfd146a IR |
1648 | if (modifier == NARROW) |
1649 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
1650 | else | |
1651 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1652 | ||
1653 | /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies | |
1654 | this, so we can safely override NCOPIES with 1 here. */ | |
1655 | if (slp_node) | |
1656 | ncopies = 1; | |
b8698a0f | 1657 | |
ebfd146a IR |
1658 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
1659 | needs to be generated. */ | |
1660 | gcc_assert (ncopies >= 1); | |
1661 | ||
ebfd146a IR |
1662 | /* Supportable by target? */ |
1663 | if ((modifier == NONE | |
88dd7150 | 1664 | && !targetm.vectorize.builtin_conversion (code, vectype_out, vectype_in)) |
ebfd146a | 1665 | || (modifier == WIDEN |
b690cc0f RG |
1666 | && !supportable_widening_operation (code, stmt, |
1667 | vectype_out, vectype_in, | |
ebfd146a IR |
1668 | &decl1, &decl2, |
1669 | &code1, &code2, | |
1670 | &dummy_int, &dummy)) | |
1671 | || (modifier == NARROW | |
b690cc0f | 1672 | && !supportable_narrowing_operation (code, vectype_out, vectype_in, |
ebfd146a IR |
1673 | &code1, &dummy_int, &dummy))) |
1674 | { | |
1675 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1676 | fprintf (vect_dump, "conversion not supported by target."); | |
1677 | return false; | |
1678 | } | |
1679 | ||
1680 | if (modifier != NONE) | |
1681 | { | |
ebfd146a IR |
1682 | /* FORNOW: SLP not supported. */ |
1683 | if (STMT_SLP_TYPE (stmt_info)) | |
b8698a0f | 1684 | return false; |
ebfd146a IR |
1685 | } |
1686 | ||
1687 | if (!vec_stmt) /* transformation not required. */ | |
1688 | { | |
1689 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
1690 | return true; | |
1691 | } | |
1692 | ||
1693 | /** Transform. **/ | |
1694 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1695 | fprintf (vect_dump, "transform conversion."); | |
1696 | ||
1697 | /* Handle def. */ | |
1698 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1699 | ||
1700 | if (modifier == NONE && !slp_node) | |
1701 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
1702 | ||
1703 | prev_stmt_info = NULL; | |
1704 | switch (modifier) | |
1705 | { | |
1706 | case NONE: | |
1707 | for (j = 0; j < ncopies; j++) | |
1708 | { | |
ebfd146a | 1709 | if (j == 0) |
b8698a0f | 1710 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
1711 | else |
1712 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
1713 | ||
1714 | builtin_decl = | |
88dd7150 RG |
1715 | targetm.vectorize.builtin_conversion (code, |
1716 | vectype_out, vectype_in); | |
ebfd146a | 1717 | for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vop0); i++) |
b8698a0f | 1718 | { |
ebfd146a IR |
1719 | /* Arguments are ready. create the new vector stmt. */ |
1720 | new_stmt = gimple_build_call (builtin_decl, 1, vop0); | |
1721 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1722 | gimple_call_set_lhs (new_stmt, new_temp); | |
1723 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a IR |
1724 | if (slp_node) |
1725 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
1726 | } | |
1727 | ||
1728 | if (j == 0) | |
1729 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1730 | else | |
1731 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1732 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1733 | } | |
1734 | break; | |
1735 | ||
1736 | case WIDEN: | |
1737 | /* In case the vectorization factor (VF) is bigger than the number | |
1738 | of elements that we can fit in a vectype (nunits), we have to | |
1739 | generate more than one vector stmt - i.e - we need to "unroll" | |
1740 | the vector stmt by a factor VF/nunits. */ | |
1741 | for (j = 0; j < ncopies; j++) | |
1742 | { | |
1743 | if (j == 0) | |
1744 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
1745 | else | |
1746 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1747 | ||
ebfd146a IR |
1748 | /* Generate first half of the widened result: */ |
1749 | new_stmt | |
b8698a0f | 1750 | = vect_gen_widened_results_half (code1, decl1, |
ebfd146a IR |
1751 | vec_oprnd0, vec_oprnd1, |
1752 | unary_op, vec_dest, gsi, stmt); | |
1753 | if (j == 0) | |
1754 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1755 | else | |
1756 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1757 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1758 | ||
1759 | /* Generate second half of the widened result: */ | |
1760 | new_stmt | |
1761 | = vect_gen_widened_results_half (code2, decl2, | |
1762 | vec_oprnd0, vec_oprnd1, | |
1763 | unary_op, vec_dest, gsi, stmt); | |
1764 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1765 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1766 | } | |
1767 | break; | |
1768 | ||
1769 | case NARROW: | |
1770 | /* In case the vectorization factor (VF) is bigger than the number | |
1771 | of elements that we can fit in a vectype (nunits), we have to | |
1772 | generate more than one vector stmt - i.e - we need to "unroll" | |
1773 | the vector stmt by a factor VF/nunits. */ | |
1774 | for (j = 0; j < ncopies; j++) | |
1775 | { | |
1776 | /* Handle uses. */ | |
1777 | if (j == 0) | |
1778 | { | |
1779 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
1780 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1781 | } | |
1782 | else | |
1783 | { | |
1784 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1); | |
1785 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
1786 | } | |
1787 | ||
1788 | /* Arguments are ready. Create the new vector stmt. */ | |
ebfd146a IR |
1789 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, vec_oprnd0, |
1790 | vec_oprnd1); | |
1791 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1792 | gimple_assign_set_lhs (new_stmt, new_temp); | |
1793 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1794 | ||
1795 | if (j == 0) | |
1796 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
1797 | else | |
1798 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1799 | ||
1800 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1801 | } | |
1802 | ||
1803 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
1804 | } | |
1805 | ||
1806 | if (vec_oprnds0) | |
b8698a0f | 1807 | VEC_free (tree, heap, vec_oprnds0); |
ebfd146a IR |
1808 | |
1809 | return true; | |
1810 | } | |
1811 | /* Function vectorizable_assignment. | |
1812 | ||
b8698a0f L |
1813 | Check if STMT performs an assignment (copy) that can be vectorized. |
1814 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1815 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1816 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1817 | ||
1818 | static bool | |
1819 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
1820 | gimple *vec_stmt, slp_tree slp_node) | |
1821 | { | |
1822 | tree vec_dest; | |
1823 | tree scalar_dest; | |
1824 | tree op; | |
1825 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1826 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1827 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1828 | tree new_temp; | |
1829 | tree def; | |
1830 | gimple def_stmt; | |
1831 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
1832 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
1833 | int ncopies; | |
f18b55bd | 1834 | int i, j; |
ebfd146a IR |
1835 | VEC(tree,heap) *vec_oprnds = NULL; |
1836 | tree vop; | |
a70d6342 | 1837 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
f18b55bd IR |
1838 | gimple new_stmt = NULL; |
1839 | stmt_vec_info prev_stmt_info = NULL; | |
ebfd146a IR |
1840 | |
1841 | /* Multiple types in SLP are handled by creating the appropriate number of | |
1842 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
1843 | case of SLP. */ | |
1844 | if (slp_node) | |
1845 | ncopies = 1; | |
1846 | else | |
1847 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1848 | ||
1849 | gcc_assert (ncopies >= 1); | |
ebfd146a | 1850 | |
a70d6342 | 1851 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1852 | return false; |
1853 | ||
8644a673 | 1854 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1855 | return false; |
1856 | ||
1857 | /* Is vectorizable assignment? */ | |
1858 | if (!is_gimple_assign (stmt)) | |
1859 | return false; | |
1860 | ||
1861 | scalar_dest = gimple_assign_lhs (stmt); | |
1862 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
1863 | return false; | |
1864 | ||
1865 | if (gimple_assign_single_p (stmt) | |
1866 | || gimple_assign_rhs_code (stmt) == PAREN_EXPR) | |
1867 | op = gimple_assign_rhs1 (stmt); | |
1868 | else | |
1869 | return false; | |
1870 | ||
a70d6342 | 1871 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt[0])) |
ebfd146a IR |
1872 | { |
1873 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1874 | fprintf (vect_dump, "use not simple."); | |
1875 | return false; | |
1876 | } | |
1877 | ||
1878 | if (!vec_stmt) /* transformation not required. */ | |
1879 | { | |
1880 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
1881 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1882 | fprintf (vect_dump, "=== vectorizable_assignment ==="); | |
1883 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
1884 | return true; | |
1885 | } | |
1886 | ||
1887 | /** Transform. **/ | |
1888 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1889 | fprintf (vect_dump, "transform assignment."); | |
1890 | ||
1891 | /* Handle def. */ | |
1892 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
1893 | ||
1894 | /* Handle use. */ | |
f18b55bd | 1895 | for (j = 0; j < ncopies; j++) |
ebfd146a | 1896 | { |
f18b55bd IR |
1897 | /* Handle uses. */ |
1898 | if (j == 0) | |
1899 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); | |
1900 | else | |
1901 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
1902 | ||
1903 | /* Arguments are ready. create the new vector stmt. */ | |
1904 | for (i = 0; VEC_iterate (tree, vec_oprnds, i, vop); i++) | |
1905 | { | |
1906 | new_stmt = gimple_build_assign (vec_dest, vop); | |
1907 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1908 | gimple_assign_set_lhs (new_stmt, new_temp); | |
1909 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1910 | if (slp_node) | |
1911 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
1912 | } | |
ebfd146a IR |
1913 | |
1914 | if (slp_node) | |
f18b55bd IR |
1915 | continue; |
1916 | ||
1917 | if (j == 0) | |
1918 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1919 | else | |
1920 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1921 | ||
1922 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1923 | } | |
b8698a0f L |
1924 | |
1925 | VEC_free (tree, heap, vec_oprnds); | |
ebfd146a IR |
1926 | return true; |
1927 | } | |
1928 | ||
1929 | /* Function vectorizable_operation. | |
1930 | ||
b8698a0f L |
1931 | Check if STMT performs a binary or unary operation that can be vectorized. |
1932 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1933 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1934 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1935 | ||
1936 | static bool | |
1937 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
1938 | gimple *vec_stmt, slp_tree slp_node) | |
1939 | { | |
1940 | tree vec_dest; | |
1941 | tree scalar_dest; | |
1942 | tree op0, op1 = NULL; | |
1943 | tree vec_oprnd1 = NULL_TREE; | |
1944 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b690cc0f | 1945 | tree vectype; |
ebfd146a IR |
1946 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
1947 | enum tree_code code; | |
1948 | enum machine_mode vec_mode; | |
1949 | tree new_temp; | |
1950 | int op_type; | |
1951 | optab optab; | |
1952 | int icode; | |
1953 | enum machine_mode optab_op2_mode; | |
1954 | tree def; | |
1955 | gimple def_stmt; | |
1956 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
1957 | gimple new_stmt = NULL; | |
1958 | stmt_vec_info prev_stmt_info; | |
b690cc0f | 1959 | int nunits_in; |
ebfd146a IR |
1960 | int nunits_out; |
1961 | tree vectype_out; | |
1962 | int ncopies; | |
1963 | int j, i; | |
1964 | VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
1965 | tree vop0, vop1; | |
1966 | unsigned int k; | |
ebfd146a | 1967 | bool scalar_shift_arg = false; |
a70d6342 IR |
1968 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
1969 | int vf; | |
1970 | ||
a70d6342 | 1971 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1972 | return false; |
1973 | ||
8644a673 | 1974 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1975 | return false; |
1976 | ||
1977 | /* Is STMT a vectorizable binary/unary operation? */ | |
1978 | if (!is_gimple_assign (stmt)) | |
1979 | return false; | |
1980 | ||
1981 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
1982 | return false; | |
1983 | ||
ebfd146a IR |
1984 | code = gimple_assign_rhs_code (stmt); |
1985 | ||
1986 | /* For pointer addition, we should use the normal plus for | |
1987 | the vector addition. */ | |
1988 | if (code == POINTER_PLUS_EXPR) | |
1989 | code = PLUS_EXPR; | |
1990 | ||
1991 | /* Support only unary or binary operations. */ | |
1992 | op_type = TREE_CODE_LENGTH (code); | |
1993 | if (op_type != unary_op && op_type != binary_op) | |
1994 | { | |
1995 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1996 | fprintf (vect_dump, "num. args = %d (not unary/binary op).", op_type); | |
1997 | return false; | |
1998 | } | |
1999 | ||
b690cc0f RG |
2000 | scalar_dest = gimple_assign_lhs (stmt); |
2001 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2002 | ||
ebfd146a | 2003 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
2004 | if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo, |
2005 | &def_stmt, &def, &dt[0], &vectype)) | |
ebfd146a IR |
2006 | { |
2007 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2008 | fprintf (vect_dump, "use not simple."); | |
2009 | return false; | |
2010 | } | |
b690cc0f RG |
2011 | /* If op0 is an external or constant def use a vector type with |
2012 | the same size as the output vector type. */ | |
2013 | if (!vectype) | |
2014 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
2015 | if (vec_stmt) |
2016 | gcc_assert (vectype); | |
2017 | if (!vectype) | |
2018 | { | |
2019 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2020 | { | |
2021 | fprintf (vect_dump, "no vectype for scalar type "); | |
2022 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2023 | } | |
2024 | ||
2025 | return false; | |
2026 | } | |
b690cc0f RG |
2027 | |
2028 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
2029 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
2030 | if (nunits_out != nunits_in) | |
2031 | return false; | |
ebfd146a IR |
2032 | |
2033 | if (op_type == binary_op) | |
2034 | { | |
2035 | op1 = gimple_assign_rhs2 (stmt); | |
b8698a0f | 2036 | if (!vect_is_simple_use (op1, loop_vinfo, bb_vinfo, &def_stmt, &def, |
a70d6342 | 2037 | &dt[1])) |
ebfd146a IR |
2038 | { |
2039 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2040 | fprintf (vect_dump, "use not simple."); | |
2041 | return false; | |
2042 | } | |
2043 | } | |
2044 | ||
b690cc0f RG |
2045 | if (loop_vinfo) |
2046 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
2047 | else | |
2048 | vf = 1; | |
2049 | ||
2050 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2051 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2052 | case of SLP. */ | |
2053 | if (slp_node) | |
2054 | ncopies = 1; | |
2055 | else | |
2056 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2057 | ||
2058 | gcc_assert (ncopies >= 1); | |
2059 | ||
ebfd146a IR |
2060 | /* If this is a shift/rotate, determine whether the shift amount is a vector, |
2061 | or scalar. If the shift/rotate amount is a vector, use the vector/vector | |
2062 | shift optabs. */ | |
2063 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
2064 | || code == RROTATE_EXPR) | |
2065 | { | |
ebfd146a | 2066 | /* vector shifted by vector */ |
8644a673 | 2067 | if (dt[1] == vect_internal_def) |
ebfd146a IR |
2068 | { |
2069 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
2070 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2071 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
2072 | } | |
2073 | ||
2074 | /* See if the machine has a vector shifted by scalar insn and if not | |
2075 | then see if it has a vector shifted by vector insn */ | |
8644a673 | 2076 | else if (dt[1] == vect_constant_def || dt[1] == vect_external_def) |
ebfd146a IR |
2077 | { |
2078 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
2079 | if (optab | |
2080 | && (optab_handler (optab, TYPE_MODE (vectype))->insn_code | |
2081 | != CODE_FOR_nothing)) | |
2082 | { | |
2083 | scalar_shift_arg = true; | |
2084 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2085 | fprintf (vect_dump, "vector/scalar shift/rotate found."); | |
2086 | } | |
2087 | else | |
2088 | { | |
2089 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
ad6c0864 | 2090 | if (optab |
ebfd146a IR |
2091 | && (optab_handler (optab, TYPE_MODE (vectype))->insn_code |
2092 | != CODE_FOR_nothing)) | |
ad6c0864 MM |
2093 | { |
2094 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2095 | fprintf (vect_dump, "vector/vector shift/rotate found."); | |
2096 | ||
2097 | /* Unlike the other binary operators, shifts/rotates have | |
2098 | the rhs being int, instead of the same type as the lhs, | |
2099 | so make sure the scalar is the right type if we are | |
2100 | dealing with vectors of short/char. */ | |
2101 | if (dt[1] == vect_constant_def) | |
2102 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
2103 | } | |
ebfd146a IR |
2104 | } |
2105 | } | |
2106 | ||
2107 | else | |
2108 | { | |
2109 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2110 | fprintf (vect_dump, "operand mode requires invariant argument."); | |
2111 | return false; | |
2112 | } | |
2113 | } | |
2114 | else | |
2115 | optab = optab_for_tree_code (code, vectype, optab_default); | |
2116 | ||
2117 | /* Supportable by target? */ | |
2118 | if (!optab) | |
2119 | { | |
2120 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2121 | fprintf (vect_dump, "no optab."); | |
2122 | return false; | |
2123 | } | |
2124 | vec_mode = TYPE_MODE (vectype); | |
2125 | icode = (int) optab_handler (optab, vec_mode)->insn_code; | |
2126 | if (icode == CODE_FOR_nothing) | |
2127 | { | |
2128 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2129 | fprintf (vect_dump, "op not supported by target."); | |
2130 | /* Check only during analysis. */ | |
2131 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
a70d6342 | 2132 | || (vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
2133 | && !vec_stmt)) |
2134 | return false; | |
2135 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2136 | fprintf (vect_dump, "proceeding using word mode."); | |
2137 | } | |
2138 | ||
2139 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
2140 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
a70d6342 | 2141 | && vf < vect_min_worthwhile_factor (code) |
ebfd146a IR |
2142 | && !vec_stmt) |
2143 | { | |
2144 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2145 | fprintf (vect_dump, "not worthwhile without SIMD support."); | |
2146 | return false; | |
2147 | } | |
2148 | ||
2149 | if (!vec_stmt) /* transformation not required. */ | |
2150 | { | |
2151 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; | |
2152 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2153 | fprintf (vect_dump, "=== vectorizable_operation ==="); | |
2154 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2155 | return true; | |
2156 | } | |
2157 | ||
2158 | /** Transform. **/ | |
2159 | ||
2160 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2161 | fprintf (vect_dump, "transform binary/unary operation."); | |
2162 | ||
2163 | /* Handle def. */ | |
2164 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2165 | ||
b8698a0f | 2166 | /* Allocate VECs for vector operands. In case of SLP, vector operands are |
ebfd146a IR |
2167 | created in the previous stages of the recursion, so no allocation is |
2168 | needed, except for the case of shift with scalar shift argument. In that | |
2169 | case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to | |
2170 | be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE. | |
b8698a0f L |
2171 | In case of loop-based vectorization we allocate VECs of size 1. We |
2172 | allocate VEC_OPRNDS1 only in case of binary operation. */ | |
ebfd146a IR |
2173 | if (!slp_node) |
2174 | { | |
2175 | vec_oprnds0 = VEC_alloc (tree, heap, 1); | |
2176 | if (op_type == binary_op) | |
2177 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
2178 | } | |
2179 | else if (scalar_shift_arg) | |
b8698a0f | 2180 | vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size); |
ebfd146a IR |
2181 | |
2182 | /* In case the vectorization factor (VF) is bigger than the number | |
2183 | of elements that we can fit in a vectype (nunits), we have to generate | |
2184 | more than one vector stmt - i.e - we need to "unroll" the | |
2185 | vector stmt by a factor VF/nunits. In doing so, we record a pointer | |
2186 | from one copy of the vector stmt to the next, in the field | |
2187 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
2188 | stages to find the correct vector defs to be used when vectorizing | |
2189 | stmts that use the defs of the current stmt. The example below illustrates | |
2190 | the vectorization process when VF=16 and nunits=4 (i.e - we need to create | |
2191 | 4 vectorized stmts): | |
2192 | ||
2193 | before vectorization: | |
2194 | RELATED_STMT VEC_STMT | |
2195 | S1: x = memref - - | |
2196 | S2: z = x + 1 - - | |
2197 | ||
2198 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
2199 | there): | |
2200 | RELATED_STMT VEC_STMT | |
2201 | VS1_0: vx0 = memref0 VS1_1 - | |
2202 | VS1_1: vx1 = memref1 VS1_2 - | |
2203 | VS1_2: vx2 = memref2 VS1_3 - | |
2204 | VS1_3: vx3 = memref3 - - | |
2205 | S1: x = load - VS1_0 | |
2206 | S2: z = x + 1 - - | |
2207 | ||
2208 | step2: vectorize stmt S2 (done here): | |
2209 | To vectorize stmt S2 we first need to find the relevant vector | |
2210 | def for the first operand 'x'. This is, as usual, obtained from | |
2211 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
2212 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
2213 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
2214 | the vector stmt VS2_0, and as usual, record it in the | |
2215 | STMT_VINFO_VEC_STMT of stmt S2. | |
2216 | When creating the second copy (VS2_1), we obtain the relevant vector | |
2217 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
2218 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
2219 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
2220 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
2221 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
2222 | chain of stmts and pointers: | |
2223 | RELATED_STMT VEC_STMT | |
2224 | VS1_0: vx0 = memref0 VS1_1 - | |
2225 | VS1_1: vx1 = memref1 VS1_2 - | |
2226 | VS1_2: vx2 = memref2 VS1_3 - | |
2227 | VS1_3: vx3 = memref3 - - | |
2228 | S1: x = load - VS1_0 | |
2229 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
2230 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
2231 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
2232 | VS2_3: vz3 = vx3 + v1 - - | |
2233 | S2: z = x + 1 - VS2_0 */ | |
2234 | ||
2235 | prev_stmt_info = NULL; | |
2236 | for (j = 0; j < ncopies; j++) | |
2237 | { | |
2238 | /* Handle uses. */ | |
2239 | if (j == 0) | |
2240 | { | |
2241 | if (op_type == binary_op && scalar_shift_arg) | |
2242 | { | |
b8698a0f L |
2243 | /* Vector shl and shr insn patterns can be defined with scalar |
2244 | operand 2 (shift operand). In this case, use constant or loop | |
2245 | invariant op1 directly, without extending it to vector mode | |
ebfd146a IR |
2246 | first. */ |
2247 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
2248 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
2249 | { | |
2250 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2251 | fprintf (vect_dump, "operand 1 using scalar mode."); | |
2252 | vec_oprnd1 = op1; | |
2253 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2254 | if (slp_node) | |
2255 | { | |
2256 | /* Store vec_oprnd1 for every vector stmt to be created | |
2257 | for SLP_NODE. We check during the analysis that all the | |
b8698a0f L |
2258 | shift arguments are the same. |
2259 | TODO: Allow different constants for different vector | |
2260 | stmts generated for an SLP instance. */ | |
ebfd146a IR |
2261 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) |
2262 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2263 | } | |
2264 | } | |
2265 | } | |
b8698a0f L |
2266 | |
2267 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
2268 | (a special case for certain kind of vector shifts); otherwise, | |
ebfd146a IR |
2269 | operand 1 should be of a vector type (the usual case). */ |
2270 | if (op_type == binary_op && !vec_oprnd1) | |
b8698a0f | 2271 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, |
ebfd146a IR |
2272 | slp_node); |
2273 | else | |
b8698a0f | 2274 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, |
ebfd146a IR |
2275 | slp_node); |
2276 | } | |
2277 | else | |
2278 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
2279 | ||
2280 | /* Arguments are ready. Create the new vector stmt. */ | |
2281 | for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vop0); i++) | |
2282 | { | |
2283 | vop1 = ((op_type == binary_op) | |
2284 | ? VEC_index (tree, vec_oprnds1, i) : NULL); | |
2285 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2286 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2287 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2288 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2289 | if (slp_node) | |
2290 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2291 | } | |
2292 | ||
2293 | if (slp_node) | |
2294 | continue; | |
2295 | ||
2296 | if (j == 0) | |
2297 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2298 | else | |
2299 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2300 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2301 | } | |
2302 | ||
2303 | VEC_free (tree, heap, vec_oprnds0); | |
2304 | if (vec_oprnds1) | |
2305 | VEC_free (tree, heap, vec_oprnds1); | |
2306 | ||
2307 | return true; | |
2308 | } | |
2309 | ||
2310 | ||
2311 | /* Get vectorized definitions for loop-based vectorization. For the first | |
b8698a0f L |
2312 | operand we call vect_get_vec_def_for_operand() (with OPRND containing |
2313 | scalar operand), and for the rest we get a copy with | |
ebfd146a IR |
2314 | vect_get_vec_def_for_stmt_copy() using the previous vector definition |
2315 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
2316 | The vectors are collected into VEC_OPRNDS. */ | |
2317 | ||
2318 | static void | |
b8698a0f | 2319 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, |
ebfd146a IR |
2320 | VEC (tree, heap) **vec_oprnds, int multi_step_cvt) |
2321 | { | |
2322 | tree vec_oprnd; | |
2323 | ||
2324 | /* Get first vector operand. */ | |
2325 | /* All the vector operands except the very first one (that is scalar oprnd) | |
2326 | are stmt copies. */ | |
b8698a0f | 2327 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) |
ebfd146a IR |
2328 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); |
2329 | else | |
2330 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
2331 | ||
2332 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
2333 | ||
2334 | /* Get second vector operand. */ | |
2335 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
2336 | VEC_quick_push (tree, *vec_oprnds, vec_oprnd); | |
b8698a0f | 2337 | |
ebfd146a IR |
2338 | *oprnd = vec_oprnd; |
2339 | ||
b8698a0f | 2340 | /* For conversion in multiple steps, continue to get operands |
ebfd146a IR |
2341 | recursively. */ |
2342 | if (multi_step_cvt) | |
b8698a0f | 2343 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); |
ebfd146a IR |
2344 | } |
2345 | ||
2346 | ||
2347 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
b8698a0f | 2348 | For multi-step conversions store the resulting vectors and call the function |
ebfd146a IR |
2349 | recursively. */ |
2350 | ||
2351 | static void | |
2352 | vect_create_vectorized_demotion_stmts (VEC (tree, heap) **vec_oprnds, | |
2353 | int multi_step_cvt, gimple stmt, | |
2354 | VEC (tree, heap) *vec_dsts, | |
2355 | gimple_stmt_iterator *gsi, | |
2356 | slp_tree slp_node, enum tree_code code, | |
2357 | stmt_vec_info *prev_stmt_info) | |
2358 | { | |
2359 | unsigned int i; | |
2360 | tree vop0, vop1, new_tmp, vec_dest; | |
2361 | gimple new_stmt; | |
2362 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2363 | ||
b8698a0f | 2364 | vec_dest = VEC_pop (tree, vec_dsts); |
ebfd146a IR |
2365 | |
2366 | for (i = 0; i < VEC_length (tree, *vec_oprnds); i += 2) | |
2367 | { | |
2368 | /* Create demotion operation. */ | |
2369 | vop0 = VEC_index (tree, *vec_oprnds, i); | |
2370 | vop1 = VEC_index (tree, *vec_oprnds, i + 1); | |
2371 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); | |
2372 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
2373 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
2374 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2375 | ||
2376 | if (multi_step_cvt) | |
2377 | /* Store the resulting vector for next recursive call. */ | |
b8698a0f | 2378 | VEC_replace (tree, *vec_oprnds, i/2, new_tmp); |
ebfd146a IR |
2379 | else |
2380 | { | |
b8698a0f | 2381 | /* This is the last step of the conversion sequence. Store the |
ebfd146a IR |
2382 | vectors in SLP_NODE or in vector info of the scalar statement |
2383 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
2384 | if (slp_node) | |
2385 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
2386 | else | |
2387 | { | |
2388 | if (!*prev_stmt_info) | |
2389 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2390 | else | |
2391 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
2392 | ||
2393 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2394 | } | |
2395 | } | |
2396 | } | |
2397 | ||
2398 | /* For multi-step demotion operations we first generate demotion operations | |
b8698a0f | 2399 | from the source type to the intermediate types, and then combine the |
ebfd146a IR |
2400 | results (stored in VEC_OPRNDS) in demotion operation to the destination |
2401 | type. */ | |
2402 | if (multi_step_cvt) | |
2403 | { | |
2404 | /* At each level of recursion we have have of the operands we had at the | |
2405 | previous level. */ | |
2406 | VEC_truncate (tree, *vec_oprnds, (i+1)/2); | |
b8698a0f | 2407 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
ebfd146a IR |
2408 | stmt, vec_dsts, gsi, slp_node, |
2409 | code, prev_stmt_info); | |
2410 | } | |
2411 | } | |
2412 | ||
2413 | ||
2414 | /* Function vectorizable_type_demotion | |
2415 | ||
2416 | Check if STMT performs a binary or unary operation that involves | |
2417 | type demotion, and if it can be vectorized. | |
2418 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2419 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2420 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2421 | ||
2422 | static bool | |
2423 | vectorizable_type_demotion (gimple stmt, gimple_stmt_iterator *gsi, | |
2424 | gimple *vec_stmt, slp_tree slp_node) | |
2425 | { | |
2426 | tree vec_dest; | |
2427 | tree scalar_dest; | |
2428 | tree op0; | |
2429 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2430 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2431 | enum tree_code code, code1 = ERROR_MARK; | |
2432 | tree def; | |
2433 | gimple def_stmt; | |
2434 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2435 | stmt_vec_info prev_stmt_info; | |
2436 | int nunits_in; | |
2437 | int nunits_out; | |
2438 | tree vectype_out; | |
2439 | int ncopies; | |
2440 | int j, i; | |
2441 | tree vectype_in; | |
2442 | int multi_step_cvt = 0; | |
2443 | VEC (tree, heap) *vec_oprnds0 = NULL; | |
2444 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL; | |
2445 | tree last_oprnd, intermediate_type; | |
2446 | ||
a70d6342 IR |
2447 | /* FORNOW: not supported by basic block SLP vectorization. */ |
2448 | gcc_assert (loop_vinfo); | |
2449 | ||
ebfd146a IR |
2450 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
2451 | return false; | |
2452 | ||
8644a673 | 2453 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2454 | return false; |
2455 | ||
2456 | /* Is STMT a vectorizable type-demotion operation? */ | |
2457 | if (!is_gimple_assign (stmt)) | |
2458 | return false; | |
2459 | ||
2460 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2461 | return false; | |
2462 | ||
2463 | code = gimple_assign_rhs_code (stmt); | |
2464 | if (!CONVERT_EXPR_CODE_P (code)) | |
2465 | return false; | |
2466 | ||
b690cc0f RG |
2467 | scalar_dest = gimple_assign_lhs (stmt); |
2468 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2469 | ||
2470 | /* Check the operands of the operation. */ | |
ebfd146a | 2471 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
2472 | if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) |
2473 | && INTEGRAL_TYPE_P (TREE_TYPE (op0))) | |
2474 | || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest)) | |
2475 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0)) | |
2476 | && CONVERT_EXPR_CODE_P (code)))) | |
2477 | return false; | |
2478 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
2479 | &def_stmt, &def, &dt[0], &vectype_in)) | |
2480 | { | |
2481 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2482 | fprintf (vect_dump, "use not simple."); | |
2483 | return false; | |
2484 | } | |
2485 | /* If op0 is an external def use a vector type with the | |
2486 | same size as the output vector type if possible. */ | |
2487 | if (!vectype_in) | |
2488 | vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
2489 | if (vec_stmt) |
2490 | gcc_assert (vectype_in); | |
ebfd146a | 2491 | if (!vectype_in) |
7d8930a0 IR |
2492 | { |
2493 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2494 | { | |
2495 | fprintf (vect_dump, "no vectype for scalar type "); | |
2496 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2497 | } | |
2498 | ||
2499 | return false; | |
2500 | } | |
ebfd146a | 2501 | |
b690cc0f | 2502 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
ebfd146a IR |
2503 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); |
2504 | if (nunits_in >= nunits_out) | |
2505 | return false; | |
2506 | ||
2507 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2508 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2509 | case of SLP. */ | |
2510 | if (slp_node) | |
2511 | ncopies = 1; | |
2512 | else | |
2513 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
ebfd146a IR |
2514 | gcc_assert (ncopies >= 1); |
2515 | ||
ebfd146a | 2516 | /* Supportable by target? */ |
b690cc0f RG |
2517 | if (!supportable_narrowing_operation (code, vectype_out, vectype_in, |
2518 | &code1, &multi_step_cvt, &interm_types)) | |
ebfd146a IR |
2519 | return false; |
2520 | ||
ebfd146a IR |
2521 | if (!vec_stmt) /* transformation not required. */ |
2522 | { | |
2523 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
2524 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2525 | fprintf (vect_dump, "=== vectorizable_demotion ==="); | |
2526 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL); | |
2527 | return true; | |
2528 | } | |
2529 | ||
2530 | /** Transform. **/ | |
2531 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2532 | fprintf (vect_dump, "transform type demotion operation. ncopies = %d.", | |
2533 | ncopies); | |
2534 | ||
b8698a0f L |
2535 | /* In case of multi-step demotion, we first generate demotion operations to |
2536 | the intermediate types, and then from that types to the final one. | |
ebfd146a | 2537 | We create vector destinations for the intermediate type (TYPES) received |
b8698a0f | 2538 | from supportable_narrowing_operation, and store them in the correct order |
ebfd146a IR |
2539 | for future use in vect_create_vectorized_demotion_stmts(). */ |
2540 | if (multi_step_cvt) | |
2541 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
2542 | else | |
2543 | vec_dsts = VEC_alloc (tree, heap, 1); | |
b8698a0f | 2544 | |
ebfd146a IR |
2545 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); |
2546 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2547 | ||
2548 | if (multi_step_cvt) | |
2549 | { | |
b8698a0f | 2550 | for (i = VEC_length (tree, interm_types) - 1; |
ebfd146a IR |
2551 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) |
2552 | { | |
b8698a0f | 2553 | vec_dest = vect_create_destination_var (scalar_dest, |
ebfd146a IR |
2554 | intermediate_type); |
2555 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2556 | } | |
2557 | } | |
2558 | ||
2559 | /* In case the vectorization factor (VF) is bigger than the number | |
2560 | of elements that we can fit in a vectype (nunits), we have to generate | |
2561 | more than one vector stmt - i.e - we need to "unroll" the | |
2562 | vector stmt by a factor VF/nunits. */ | |
2563 | last_oprnd = op0; | |
2564 | prev_stmt_info = NULL; | |
2565 | for (j = 0; j < ncopies; j++) | |
2566 | { | |
2567 | /* Handle uses. */ | |
2568 | if (slp_node) | |
b5aeb3bb | 2569 | vect_get_slp_defs (slp_node, &vec_oprnds0, NULL, -1); |
ebfd146a IR |
2570 | else |
2571 | { | |
2572 | VEC_free (tree, heap, vec_oprnds0); | |
2573 | vec_oprnds0 = VEC_alloc (tree, heap, | |
2574 | (multi_step_cvt ? vect_pow2 (multi_step_cvt) * 2 : 2)); | |
b8698a0f | 2575 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
ebfd146a IR |
2576 | vect_pow2 (multi_step_cvt) - 1); |
2577 | } | |
2578 | ||
2579 | /* Arguments are ready. Create the new vector stmts. */ | |
2580 | tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts); | |
b8698a0f | 2581 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, |
ebfd146a | 2582 | multi_step_cvt, stmt, tmp_vec_dsts, |
b8698a0f | 2583 | gsi, slp_node, code1, |
ebfd146a IR |
2584 | &prev_stmt_info); |
2585 | } | |
2586 | ||
2587 | VEC_free (tree, heap, vec_oprnds0); | |
2588 | VEC_free (tree, heap, vec_dsts); | |
2589 | VEC_free (tree, heap, tmp_vec_dsts); | |
2590 | VEC_free (tree, heap, interm_types); | |
2591 | ||
2592 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2593 | return true; | |
2594 | } | |
2595 | ||
2596 | ||
2597 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
b8698a0f | 2598 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store |
ebfd146a IR |
2599 | the resulting vectors and call the function recursively. */ |
2600 | ||
2601 | static void | |
2602 | vect_create_vectorized_promotion_stmts (VEC (tree, heap) **vec_oprnds0, | |
2603 | VEC (tree, heap) **vec_oprnds1, | |
2604 | int multi_step_cvt, gimple stmt, | |
2605 | VEC (tree, heap) *vec_dsts, | |
2606 | gimple_stmt_iterator *gsi, | |
2607 | slp_tree slp_node, enum tree_code code1, | |
b8698a0f | 2608 | enum tree_code code2, tree decl1, |
ebfd146a IR |
2609 | tree decl2, int op_type, |
2610 | stmt_vec_info *prev_stmt_info) | |
2611 | { | |
2612 | int i; | |
2613 | tree vop0, vop1, new_tmp1, new_tmp2, vec_dest; | |
2614 | gimple new_stmt1, new_stmt2; | |
2615 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2616 | VEC (tree, heap) *vec_tmp; | |
2617 | ||
2618 | vec_dest = VEC_pop (tree, vec_dsts); | |
2619 | vec_tmp = VEC_alloc (tree, heap, VEC_length (tree, *vec_oprnds0) * 2); | |
2620 | ||
2621 | for (i = 0; VEC_iterate (tree, *vec_oprnds0, i, vop0); i++) | |
2622 | { | |
2623 | if (op_type == binary_op) | |
2624 | vop1 = VEC_index (tree, *vec_oprnds1, i); | |
2625 | else | |
2626 | vop1 = NULL_TREE; | |
2627 | ||
2628 | /* Generate the two halves of promotion operation. */ | |
b8698a0f | 2629 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, |
ebfd146a IR |
2630 | op_type, vec_dest, gsi, stmt); |
2631 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
2632 | op_type, vec_dest, gsi, stmt); | |
2633 | if (is_gimple_call (new_stmt1)) | |
2634 | { | |
2635 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
2636 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
2637 | } | |
2638 | else | |
2639 | { | |
2640 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
2641 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
2642 | } | |
2643 | ||
2644 | if (multi_step_cvt) | |
2645 | { | |
2646 | /* Store the results for the recursive call. */ | |
2647 | VEC_quick_push (tree, vec_tmp, new_tmp1); | |
2648 | VEC_quick_push (tree, vec_tmp, new_tmp2); | |
2649 | } | |
2650 | else | |
2651 | { | |
2652 | /* Last step of promotion sequience - store the results. */ | |
2653 | if (slp_node) | |
2654 | { | |
2655 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt1); | |
2656 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt2); | |
2657 | } | |
2658 | else | |
2659 | { | |
2660 | if (!*prev_stmt_info) | |
2661 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt1; | |
2662 | else | |
2663 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt1; | |
2664 | ||
2665 | *prev_stmt_info = vinfo_for_stmt (new_stmt1); | |
2666 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt2; | |
2667 | *prev_stmt_info = vinfo_for_stmt (new_stmt2); | |
2668 | } | |
2669 | } | |
2670 | } | |
2671 | ||
2672 | if (multi_step_cvt) | |
2673 | { | |
b8698a0f | 2674 | /* For multi-step promotion operation we first generate we call the |
ebfd146a IR |
2675 | function recurcively for every stage. We start from the input type, |
2676 | create promotion operations to the intermediate types, and then | |
2677 | create promotions to the output type. */ | |
2678 | *vec_oprnds0 = VEC_copy (tree, heap, vec_tmp); | |
2679 | VEC_free (tree, heap, vec_tmp); | |
2680 | vect_create_vectorized_promotion_stmts (vec_oprnds0, vec_oprnds1, | |
2681 | multi_step_cvt - 1, stmt, | |
2682 | vec_dsts, gsi, slp_node, code1, | |
2683 | code2, decl2, decl2, op_type, | |
2684 | prev_stmt_info); | |
2685 | } | |
2686 | } | |
b8698a0f | 2687 | |
ebfd146a IR |
2688 | |
2689 | /* Function vectorizable_type_promotion | |
2690 | ||
2691 | Check if STMT performs a binary or unary operation that involves | |
2692 | type promotion, and if it can be vectorized. | |
2693 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2694 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2695 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2696 | ||
2697 | static bool | |
2698 | vectorizable_type_promotion (gimple stmt, gimple_stmt_iterator *gsi, | |
2699 | gimple *vec_stmt, slp_tree slp_node) | |
2700 | { | |
2701 | tree vec_dest; | |
2702 | tree scalar_dest; | |
2703 | tree op0, op1 = NULL; | |
2704 | tree vec_oprnd0=NULL, vec_oprnd1=NULL; | |
2705 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2706 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2707 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
2708 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; | |
b8698a0f | 2709 | int op_type; |
ebfd146a IR |
2710 | tree def; |
2711 | gimple def_stmt; | |
2712 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2713 | stmt_vec_info prev_stmt_info; | |
2714 | int nunits_in; | |
2715 | int nunits_out; | |
2716 | tree vectype_out; | |
2717 | int ncopies; | |
2718 | int j, i; | |
2719 | tree vectype_in; | |
2720 | tree intermediate_type = NULL_TREE; | |
2721 | int multi_step_cvt = 0; | |
2722 | VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL; | |
2723 | VEC (tree, heap) *vec_dsts = NULL, *interm_types = NULL, *tmp_vec_dsts = NULL; | |
b8698a0f | 2724 | |
a70d6342 IR |
2725 | /* FORNOW: not supported by basic block SLP vectorization. */ |
2726 | gcc_assert (loop_vinfo); | |
b8698a0f | 2727 | |
ebfd146a IR |
2728 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) |
2729 | return false; | |
2730 | ||
8644a673 | 2731 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2732 | return false; |
2733 | ||
2734 | /* Is STMT a vectorizable type-promotion operation? */ | |
2735 | if (!is_gimple_assign (stmt)) | |
2736 | return false; | |
2737 | ||
2738 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2739 | return false; | |
2740 | ||
2741 | code = gimple_assign_rhs_code (stmt); | |
2742 | if (!CONVERT_EXPR_CODE_P (code) | |
2743 | && code != WIDEN_MULT_EXPR) | |
2744 | return false; | |
2745 | ||
b690cc0f RG |
2746 | scalar_dest = gimple_assign_lhs (stmt); |
2747 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2748 | ||
2749 | /* Check the operands of the operation. */ | |
ebfd146a | 2750 | op0 = gimple_assign_rhs1 (stmt); |
b690cc0f RG |
2751 | if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) |
2752 | && INTEGRAL_TYPE_P (TREE_TYPE (op0))) | |
2753 | || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest)) | |
2754 | && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0)) | |
2755 | && CONVERT_EXPR_CODE_P (code)))) | |
2756 | return false; | |
2757 | if (!vect_is_simple_use_1 (op0, loop_vinfo, NULL, | |
2758 | &def_stmt, &def, &dt[0], &vectype_in)) | |
2759 | { | |
2760 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2761 | fprintf (vect_dump, "use not simple."); | |
2762 | return false; | |
2763 | } | |
2764 | /* If op0 is an external or constant def use a vector type with | |
2765 | the same size as the output vector type. */ | |
2766 | if (!vectype_in) | |
2767 | vectype_in = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
2768 | if (vec_stmt) |
2769 | gcc_assert (vectype_in); | |
ebfd146a | 2770 | if (!vectype_in) |
7d8930a0 IR |
2771 | { |
2772 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2773 | { | |
2774 | fprintf (vect_dump, "no vectype for scalar type "); | |
2775 | print_generic_expr (vect_dump, TREE_TYPE (op0), TDF_SLIM); | |
2776 | } | |
2777 | ||
2778 | return false; | |
2779 | } | |
ebfd146a | 2780 | |
b690cc0f | 2781 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
ebfd146a IR |
2782 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); |
2783 | if (nunits_in <= nunits_out) | |
2784 | return false; | |
2785 | ||
2786 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2787 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2788 | case of SLP. */ | |
2789 | if (slp_node) | |
2790 | ncopies = 1; | |
2791 | else | |
2792 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2793 | ||
2794 | gcc_assert (ncopies >= 1); | |
2795 | ||
ebfd146a IR |
2796 | op_type = TREE_CODE_LENGTH (code); |
2797 | if (op_type == binary_op) | |
2798 | { | |
2799 | op1 = gimple_assign_rhs2 (stmt); | |
a70d6342 | 2800 | if (!vect_is_simple_use (op1, loop_vinfo, NULL, &def_stmt, &def, &dt[1])) |
ebfd146a IR |
2801 | { |
2802 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2803 | fprintf (vect_dump, "use not simple."); | |
2804 | return false; | |
2805 | } | |
2806 | } | |
2807 | ||
2808 | /* Supportable by target? */ | |
b690cc0f | 2809 | if (!supportable_widening_operation (code, stmt, vectype_out, vectype_in, |
ebfd146a IR |
2810 | &decl1, &decl2, &code1, &code2, |
2811 | &multi_step_cvt, &interm_types)) | |
2812 | return false; | |
2813 | ||
2814 | /* Binary widening operation can only be supported directly by the | |
2815 | architecture. */ | |
2816 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
2817 | ||
ebfd146a IR |
2818 | if (!vec_stmt) /* transformation not required. */ |
2819 | { | |
2820 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
2821 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2822 | fprintf (vect_dump, "=== vectorizable_promotion ==="); | |
2823 | vect_model_simple_cost (stmt_info, 2*ncopies, dt, NULL); | |
2824 | return true; | |
2825 | } | |
2826 | ||
2827 | /** Transform. **/ | |
2828 | ||
2829 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2830 | fprintf (vect_dump, "transform type promotion operation. ncopies = %d.", | |
2831 | ncopies); | |
2832 | ||
2833 | /* Handle def. */ | |
b8698a0f | 2834 | /* In case of multi-step promotion, we first generate promotion operations |
ebfd146a | 2835 | to the intermediate types, and then from that types to the final one. |
b8698a0f L |
2836 | We store vector destination in VEC_DSTS in the correct order for |
2837 | recursive creation of promotion operations in | |
ebfd146a IR |
2838 | vect_create_vectorized_promotion_stmts(). Vector destinations are created |
2839 | according to TYPES recieved from supportable_widening_operation(). */ | |
2840 | if (multi_step_cvt) | |
2841 | vec_dsts = VEC_alloc (tree, heap, multi_step_cvt + 1); | |
2842 | else | |
2843 | vec_dsts = VEC_alloc (tree, heap, 1); | |
2844 | ||
2845 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2846 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2847 | ||
2848 | if (multi_step_cvt) | |
2849 | { | |
2850 | for (i = VEC_length (tree, interm_types) - 1; | |
2851 | VEC_iterate (tree, interm_types, i, intermediate_type); i--) | |
2852 | { | |
2853 | vec_dest = vect_create_destination_var (scalar_dest, | |
2854 | intermediate_type); | |
2855 | VEC_quick_push (tree, vec_dsts, vec_dest); | |
2856 | } | |
2857 | } | |
b8698a0f | 2858 | |
ebfd146a IR |
2859 | if (!slp_node) |
2860 | { | |
b8698a0f | 2861 | vec_oprnds0 = VEC_alloc (tree, heap, |
ebfd146a IR |
2862 | (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); |
2863 | if (op_type == binary_op) | |
2864 | vec_oprnds1 = VEC_alloc (tree, heap, 1); | |
2865 | } | |
2866 | ||
2867 | /* In case the vectorization factor (VF) is bigger than the number | |
2868 | of elements that we can fit in a vectype (nunits), we have to generate | |
2869 | more than one vector stmt - i.e - we need to "unroll" the | |
2870 | vector stmt by a factor VF/nunits. */ | |
2871 | ||
2872 | prev_stmt_info = NULL; | |
2873 | for (j = 0; j < ncopies; j++) | |
2874 | { | |
2875 | /* Handle uses. */ | |
2876 | if (j == 0) | |
2877 | { | |
2878 | if (slp_node) | |
b5aeb3bb | 2879 | vect_get_slp_defs (slp_node, &vec_oprnds0, &vec_oprnds1, -1); |
ebfd146a IR |
2880 | else |
2881 | { | |
2882 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
2883 | VEC_quick_push (tree, vec_oprnds0, vec_oprnd0); | |
2884 | if (op_type == binary_op) | |
2885 | { | |
2886 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL); | |
2887 | VEC_quick_push (tree, vec_oprnds1, vec_oprnd1); | |
2888 | } | |
2889 | } | |
2890 | } | |
2891 | else | |
2892 | { | |
2893 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
2894 | VEC_replace (tree, vec_oprnds0, 0, vec_oprnd0); | |
2895 | if (op_type == binary_op) | |
2896 | { | |
2897 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1); | |
2898 | VEC_replace (tree, vec_oprnds1, 0, vec_oprnd1); | |
2899 | } | |
2900 | } | |
2901 | ||
2902 | /* Arguments are ready. Create the new vector stmts. */ | |
2903 | tmp_vec_dsts = VEC_copy (tree, heap, vec_dsts); | |
2904 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, &vec_oprnds1, | |
b8698a0f | 2905 | multi_step_cvt, stmt, |
ebfd146a IR |
2906 | tmp_vec_dsts, |
2907 | gsi, slp_node, code1, code2, | |
2908 | decl1, decl2, op_type, | |
2909 | &prev_stmt_info); | |
2910 | } | |
2911 | ||
2912 | VEC_free (tree, heap, vec_dsts); | |
2913 | VEC_free (tree, heap, tmp_vec_dsts); | |
2914 | VEC_free (tree, heap, interm_types); | |
2915 | VEC_free (tree, heap, vec_oprnds0); | |
2916 | VEC_free (tree, heap, vec_oprnds1); | |
2917 | ||
2918 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2919 | return true; | |
2920 | } | |
2921 | ||
2922 | ||
2923 | /* Function vectorizable_store. | |
2924 | ||
b8698a0f L |
2925 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
2926 | can be vectorized. | |
2927 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
2928 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2929 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2930 | ||
2931 | static bool | |
2932 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
2933 | slp_tree slp_node) | |
2934 | { | |
2935 | tree scalar_dest; | |
2936 | tree data_ref; | |
2937 | tree op; | |
2938 | tree vec_oprnd = NULL_TREE; | |
2939 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2940 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
2941 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2942 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 2943 | struct loop *loop = NULL; |
ebfd146a IR |
2944 | enum machine_mode vec_mode; |
2945 | tree dummy; | |
2946 | enum dr_alignment_support alignment_support_scheme; | |
2947 | tree def; | |
2948 | gimple def_stmt; | |
2949 | enum vect_def_type dt; | |
2950 | stmt_vec_info prev_stmt_info = NULL; | |
2951 | tree dataref_ptr = NULL_TREE; | |
2952 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2953 | int ncopies; | |
2954 | int j; | |
2955 | gimple next_stmt, first_stmt = NULL; | |
2956 | bool strided_store = false; | |
2957 | unsigned int group_size, i; | |
2958 | VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL; | |
2959 | bool inv_p; | |
2960 | VEC(tree,heap) *vec_oprnds = NULL; | |
2961 | bool slp = (slp_node != NULL); | |
ebfd146a | 2962 | unsigned int vec_num; |
a70d6342 IR |
2963 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2964 | ||
2965 | if (loop_vinfo) | |
2966 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
2967 | |
2968 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2969 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2970 | case of SLP. */ | |
2971 | if (slp) | |
2972 | ncopies = 1; | |
2973 | else | |
2974 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2975 | ||
2976 | gcc_assert (ncopies >= 1); | |
2977 | ||
2978 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 2979 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a IR |
2980 | { |
2981 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2982 | fprintf (vect_dump, "multiple types in nested loop."); | |
2983 | return false; | |
2984 | } | |
2985 | ||
a70d6342 | 2986 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2987 | return false; |
2988 | ||
8644a673 | 2989 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2990 | return false; |
2991 | ||
2992 | /* Is vectorizable store? */ | |
2993 | ||
2994 | if (!is_gimple_assign (stmt)) | |
2995 | return false; | |
2996 | ||
2997 | scalar_dest = gimple_assign_lhs (stmt); | |
2998 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
2999 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
e9dbe7bb IR |
3000 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
3001 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
3002 | && TREE_CODE (scalar_dest) != REALPART_EXPR) | |
ebfd146a IR |
3003 | return false; |
3004 | ||
3005 | gcc_assert (gimple_assign_single_p (stmt)); | |
3006 | op = gimple_assign_rhs1 (stmt); | |
a70d6342 | 3007 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, &dt)) |
ebfd146a IR |
3008 | { |
3009 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3010 | fprintf (vect_dump, "use not simple."); | |
3011 | return false; | |
3012 | } | |
3013 | ||
3014 | /* The scalar rhs type needs to be trivially convertible to the vector | |
3015 | component type. This should always be the case. */ | |
3016 | if (!useless_type_conversion_p (TREE_TYPE (vectype), TREE_TYPE (op))) | |
b8698a0f | 3017 | { |
ebfd146a IR |
3018 | if (vect_print_dump_info (REPORT_DETAILS)) |
3019 | fprintf (vect_dump, "??? operands of different types"); | |
3020 | return false; | |
3021 | } | |
3022 | ||
3023 | vec_mode = TYPE_MODE (vectype); | |
3024 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
3025 | (e.g. - array initialization with 0). */ | |
3026 | if (optab_handler (mov_optab, (int)vec_mode)->insn_code == CODE_FOR_nothing) | |
3027 | return false; | |
3028 | ||
3029 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3030 | return false; | |
3031 | ||
3032 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) | |
3033 | { | |
3034 | strided_store = true; | |
3035 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
3036 | if (!vect_strided_store_supported (vectype) | |
3037 | && !PURE_SLP_STMT (stmt_info) && !slp) | |
3038 | return false; | |
b8698a0f | 3039 | |
ebfd146a IR |
3040 | if (first_stmt == stmt) |
3041 | { | |
3042 | /* STMT is the leader of the group. Check the operands of all the | |
3043 | stmts of the group. */ | |
3044 | next_stmt = DR_GROUP_NEXT_DR (stmt_info); | |
3045 | while (next_stmt) | |
3046 | { | |
3047 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
3048 | op = gimple_assign_rhs1 (next_stmt); | |
b8698a0f | 3049 | if (!vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, |
a70d6342 | 3050 | &def, &dt)) |
ebfd146a IR |
3051 | { |
3052 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3053 | fprintf (vect_dump, "use not simple."); | |
3054 | return false; | |
3055 | } | |
3056 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); | |
3057 | } | |
3058 | } | |
3059 | } | |
3060 | ||
3061 | if (!vec_stmt) /* transformation not required. */ | |
3062 | { | |
3063 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
3064 | vect_model_store_cost (stmt_info, ncopies, dt, NULL); | |
3065 | return true; | |
3066 | } | |
3067 | ||
3068 | /** Transform. **/ | |
3069 | ||
3070 | if (strided_store) | |
3071 | { | |
3072 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3073 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
3074 | ||
3075 | DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; | |
3076 | ||
3077 | /* FORNOW */ | |
a70d6342 | 3078 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
3079 | |
3080 | /* We vectorize all the stmts of the interleaving group when we | |
3081 | reach the last stmt in the group. */ | |
b8698a0f | 3082 | if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
ebfd146a IR |
3083 | < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)) |
3084 | && !slp) | |
3085 | { | |
3086 | *vec_stmt = NULL; | |
3087 | return true; | |
3088 | } | |
3089 | ||
3090 | if (slp) | |
4b5caab7 IR |
3091 | { |
3092 | strided_store = false; | |
3093 | /* VEC_NUM is the number of vect stmts to be created for this | |
3094 | group. */ | |
3095 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
3096 | first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0); | |
3097 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3098 | } | |
ebfd146a | 3099 | else |
4b5caab7 IR |
3100 | /* VEC_NUM is the number of vect stmts to be created for this |
3101 | group. */ | |
ebfd146a IR |
3102 | vec_num = group_size; |
3103 | } | |
b8698a0f | 3104 | else |
ebfd146a IR |
3105 | { |
3106 | first_stmt = stmt; | |
3107 | first_dr = dr; | |
3108 | group_size = vec_num = 1; | |
ebfd146a | 3109 | } |
b8698a0f | 3110 | |
ebfd146a IR |
3111 | if (vect_print_dump_info (REPORT_DETAILS)) |
3112 | fprintf (vect_dump, "transform store. ncopies = %d",ncopies); | |
3113 | ||
3114 | dr_chain = VEC_alloc (tree, heap, group_size); | |
3115 | oprnds = VEC_alloc (tree, heap, group_size); | |
3116 | ||
3117 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr); | |
3118 | gcc_assert (alignment_support_scheme); | |
ebfd146a IR |
3119 | |
3120 | /* In case the vectorization factor (VF) is bigger than the number | |
3121 | of elements that we can fit in a vectype (nunits), we have to generate | |
3122 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 3123 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
3124 | vect_get_vec_def_for_copy_stmt. */ |
3125 | ||
3126 | /* In case of interleaving (non-unit strided access): | |
3127 | ||
3128 | S1: &base + 2 = x2 | |
3129 | S2: &base = x0 | |
3130 | S3: &base + 1 = x1 | |
3131 | S4: &base + 3 = x3 | |
3132 | ||
3133 | We create vectorized stores starting from base address (the access of the | |
3134 | first stmt in the chain (S2 in the above example), when the last store stmt | |
3135 | of the chain (S4) is reached: | |
3136 | ||
3137 | VS1: &base = vx2 | |
3138 | VS2: &base + vec_size*1 = vx0 | |
3139 | VS3: &base + vec_size*2 = vx1 | |
3140 | VS4: &base + vec_size*3 = vx3 | |
3141 | ||
3142 | Then permutation statements are generated: | |
3143 | ||
3144 | VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 > | |
3145 | VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 > | |
3146 | ... | |
b8698a0f | 3147 | |
ebfd146a IR |
3148 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
3149 | (the order of the data-refs in the output of vect_permute_store_chain | |
3150 | corresponds to the order of scalar stmts in the interleaving chain - see | |
3151 | the documentation of vect_permute_store_chain()). | |
3152 | ||
3153 | In case of both multiple types and interleaving, above vector stores and | |
3154 | permutation stmts are created for every copy. The result vector stmts are | |
3155 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding | |
b8698a0f | 3156 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
3157 | */ |
3158 | ||
3159 | prev_stmt_info = NULL; | |
3160 | for (j = 0; j < ncopies; j++) | |
3161 | { | |
3162 | gimple new_stmt; | |
3163 | gimple ptr_incr; | |
3164 | ||
3165 | if (j == 0) | |
3166 | { | |
3167 | if (slp) | |
3168 | { | |
3169 | /* Get vectorized arguments for SLP_NODE. */ | |
b5aeb3bb | 3170 | vect_get_slp_defs (slp_node, &vec_oprnds, NULL, -1); |
ebfd146a IR |
3171 | |
3172 | vec_oprnd = VEC_index (tree, vec_oprnds, 0); | |
3173 | } | |
3174 | else | |
3175 | { | |
b8698a0f L |
3176 | /* For interleaved stores we collect vectorized defs for all the |
3177 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
3178 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
3179 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
3180 | ||
3181 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3182 | OPRNDS are of size 1. */ | |
b8698a0f | 3183 | next_stmt = first_stmt; |
ebfd146a IR |
3184 | for (i = 0; i < group_size; i++) |
3185 | { | |
b8698a0f L |
3186 | /* Since gaps are not supported for interleaved stores, |
3187 | GROUP_SIZE is the exact number of stmts in the chain. | |
3188 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
3189 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
3190 | iteration of the loop will be executed. */ |
3191 | gcc_assert (next_stmt | |
3192 | && gimple_assign_single_p (next_stmt)); | |
3193 | op = gimple_assign_rhs1 (next_stmt); | |
3194 | ||
b8698a0f | 3195 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, |
ebfd146a | 3196 | NULL); |
b8698a0f L |
3197 | VEC_quick_push(tree, dr_chain, vec_oprnd); |
3198 | VEC_quick_push(tree, oprnds, vec_oprnd); | |
ebfd146a IR |
3199 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); |
3200 | } | |
3201 | } | |
3202 | ||
3203 | /* We should have catched mismatched types earlier. */ | |
3204 | gcc_assert (useless_type_conversion_p (vectype, | |
3205 | TREE_TYPE (vec_oprnd))); | |
b8698a0f L |
3206 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, NULL, NULL_TREE, |
3207 | &dummy, &ptr_incr, false, | |
5006671f | 3208 | &inv_p); |
a70d6342 | 3209 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 3210 | } |
b8698a0f | 3211 | else |
ebfd146a | 3212 | { |
b8698a0f L |
3213 | /* For interleaved stores we created vectorized defs for all the |
3214 | defs stored in OPRNDS in the previous iteration (previous copy). | |
3215 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
3216 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
3217 | next copy. | |
3218 | If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and | |
3219 | OPRNDS are of size 1. */ | |
3220 | for (i = 0; i < group_size; i++) | |
3221 | { | |
3222 | op = VEC_index (tree, oprnds, i); | |
b8698a0f | 3223 | vect_is_simple_use (op, loop_vinfo, bb_vinfo, &def_stmt, &def, |
a70d6342 | 3224 | &dt); |
b8698a0f | 3225 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
ebfd146a IR |
3226 | VEC_replace(tree, dr_chain, i, vec_oprnd); |
3227 | VEC_replace(tree, oprnds, i, vec_oprnd); | |
3228 | } | |
b8698a0f | 3229 | dataref_ptr = |
ebfd146a IR |
3230 | bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE); |
3231 | } | |
3232 | ||
3233 | if (strided_store) | |
3234 | { | |
b8698a0f | 3235 | result_chain = VEC_alloc (tree, heap, group_size); |
ebfd146a IR |
3236 | /* Permute. */ |
3237 | if (!vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
3238 | &result_chain)) | |
3239 | return false; | |
3240 | } | |
3241 | ||
3242 | next_stmt = first_stmt; | |
3243 | for (i = 0; i < vec_num; i++) | |
3244 | { | |
3245 | if (i > 0) | |
3246 | /* Bump the vector pointer. */ | |
3247 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
3248 | NULL_TREE); | |
3249 | ||
3250 | if (slp) | |
3251 | vec_oprnd = VEC_index (tree, vec_oprnds, i); | |
3252 | else if (strided_store) | |
b8698a0f | 3253 | /* For strided stores vectorized defs are interleaved in |
ebfd146a IR |
3254 | vect_permute_store_chain(). */ |
3255 | vec_oprnd = VEC_index (tree, result_chain, i); | |
3256 | ||
8f439681 RE |
3257 | if (aligned_access_p (first_dr)) |
3258 | data_ref = build_fold_indirect_ref (dataref_ptr); | |
3259 | else | |
3260 | { | |
3261 | int mis = DR_MISALIGNMENT (first_dr); | |
3262 | tree tmis = (mis == -1 ? size_zero_node : size_int (mis)); | |
3263 | tmis = size_binop (MULT_EXPR, tmis, size_int (BITS_PER_UNIT)); | |
3264 | data_ref = build2 (MISALIGNED_INDIRECT_REF, vectype, dataref_ptr, tmis); | |
3265 | } | |
3266 | ||
5006671f RG |
3267 | /* If accesses through a pointer to vectype do not alias the original |
3268 | memory reference we have a problem. This should never happen. */ | |
3269 | gcc_assert (alias_sets_conflict_p (get_alias_set (data_ref), | |
3270 | get_alias_set (gimple_assign_lhs (stmt)))); | |
ebfd146a IR |
3271 | |
3272 | /* Arguments are ready. Create the new vector stmt. */ | |
3273 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
3274 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3275 | mark_symbols_for_renaming (new_stmt); | |
3276 | ||
3277 | if (slp) | |
3278 | continue; | |
b8698a0f | 3279 | |
ebfd146a IR |
3280 | if (j == 0) |
3281 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3282 | else | |
3283 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3284 | ||
3285 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3286 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); | |
3287 | if (!next_stmt) | |
3288 | break; | |
3289 | } | |
3290 | } | |
3291 | ||
b8698a0f L |
3292 | VEC_free (tree, heap, dr_chain); |
3293 | VEC_free (tree, heap, oprnds); | |
ebfd146a | 3294 | if (result_chain) |
b8698a0f | 3295 | VEC_free (tree, heap, result_chain); |
ebfd146a IR |
3296 | |
3297 | return true; | |
3298 | } | |
3299 | ||
3300 | /* vectorizable_load. | |
3301 | ||
b8698a0f L |
3302 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
3303 | can be vectorized. | |
3304 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3305 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3306 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3307 | ||
3308 | static bool | |
3309 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
3310 | slp_tree slp_node, slp_instance slp_node_instance) | |
3311 | { | |
3312 | tree scalar_dest; | |
3313 | tree vec_dest = NULL; | |
3314 | tree data_ref = NULL; | |
3315 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 3316 | stmt_vec_info prev_stmt_info; |
ebfd146a | 3317 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 3318 | struct loop *loop = NULL; |
ebfd146a | 3319 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 3320 | bool nested_in_vect_loop = false; |
ebfd146a IR |
3321 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
3322 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3323 | tree new_temp; | |
3324 | int mode; | |
3325 | gimple new_stmt = NULL; | |
3326 | tree dummy; | |
3327 | enum dr_alignment_support alignment_support_scheme; | |
3328 | tree dataref_ptr = NULL_TREE; | |
3329 | gimple ptr_incr; | |
3330 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3331 | int ncopies; | |
3332 | int i, j, group_size; | |
3333 | tree msq = NULL_TREE, lsq; | |
3334 | tree offset = NULL_TREE; | |
3335 | tree realignment_token = NULL_TREE; | |
3336 | gimple phi = NULL; | |
3337 | VEC(tree,heap) *dr_chain = NULL; | |
3338 | bool strided_load = false; | |
3339 | gimple first_stmt; | |
3340 | tree scalar_type; | |
3341 | bool inv_p; | |
3342 | bool compute_in_loop = false; | |
3343 | struct loop *at_loop; | |
3344 | int vec_num; | |
3345 | bool slp = (slp_node != NULL); | |
3346 | bool slp_perm = false; | |
3347 | enum tree_code code; | |
a70d6342 IR |
3348 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3349 | int vf; | |
3350 | ||
3351 | if (loop_vinfo) | |
3352 | { | |
3353 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
3354 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
3355 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3356 | } | |
3357 | else | |
3533e503 | 3358 | vf = 1; |
ebfd146a IR |
3359 | |
3360 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3361 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3362 | case of SLP. */ | |
3363 | if (slp) | |
3364 | ncopies = 1; | |
3365 | else | |
3366 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3367 | ||
3368 | gcc_assert (ncopies >= 1); | |
3369 | ||
3370 | /* FORNOW. This restriction should be relaxed. */ | |
3371 | if (nested_in_vect_loop && ncopies > 1) | |
3372 | { | |
3373 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3374 | fprintf (vect_dump, "multiple types in nested loop."); | |
3375 | return false; | |
3376 | } | |
3377 | ||
a70d6342 | 3378 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3379 | return false; |
3380 | ||
8644a673 | 3381 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3382 | return false; |
3383 | ||
3384 | /* Is vectorizable load? */ | |
3385 | if (!is_gimple_assign (stmt)) | |
3386 | return false; | |
3387 | ||
3388 | scalar_dest = gimple_assign_lhs (stmt); | |
3389 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
3390 | return false; | |
3391 | ||
3392 | code = gimple_assign_rhs_code (stmt); | |
3393 | if (code != ARRAY_REF | |
3394 | && code != INDIRECT_REF | |
e9dbe7bb IR |
3395 | && code != COMPONENT_REF |
3396 | && code != IMAGPART_EXPR | |
3397 | && code != REALPART_EXPR) | |
ebfd146a IR |
3398 | return false; |
3399 | ||
3400 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3401 | return false; | |
3402 | ||
3403 | scalar_type = TREE_TYPE (DR_REF (dr)); | |
3404 | mode = (int) TYPE_MODE (vectype); | |
3405 | ||
3406 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
3407 | (e.g. - data copies). */ | |
3408 | if (optab_handler (mov_optab, mode)->insn_code == CODE_FOR_nothing) | |
3409 | { | |
3410 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3411 | fprintf (vect_dump, "Aligned load, but unsupported type."); | |
3412 | return false; | |
3413 | } | |
3414 | ||
3415 | /* The vector component type needs to be trivially convertible to the | |
3416 | scalar lhs. This should always be the case. */ | |
3417 | if (!useless_type_conversion_p (TREE_TYPE (scalar_dest), TREE_TYPE (vectype))) | |
b8698a0f | 3418 | { |
ebfd146a IR |
3419 | if (vect_print_dump_info (REPORT_DETAILS)) |
3420 | fprintf (vect_dump, "??? operands of different types"); | |
3421 | return false; | |
3422 | } | |
3423 | ||
3424 | /* Check if the load is a part of an interleaving chain. */ | |
3425 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) | |
3426 | { | |
3427 | strided_load = true; | |
3428 | /* FORNOW */ | |
3429 | gcc_assert (! nested_in_vect_loop); | |
3430 | ||
3431 | /* Check if interleaving is supported. */ | |
3432 | if (!vect_strided_load_supported (vectype) | |
3433 | && !PURE_SLP_STMT (stmt_info) && !slp) | |
3434 | return false; | |
3435 | } | |
3436 | ||
3437 | if (!vec_stmt) /* transformation not required. */ | |
3438 | { | |
3439 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
3440 | vect_model_load_cost (stmt_info, ncopies, NULL); | |
3441 | return true; | |
3442 | } | |
3443 | ||
3444 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3445 | fprintf (vect_dump, "transform load."); | |
3446 | ||
3447 | /** Transform. **/ | |
3448 | ||
3449 | if (strided_load) | |
3450 | { | |
3451 | first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
3452 | /* Check if the chain of loads is already vectorized. */ | |
3453 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt))) | |
3454 | { | |
3455 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
3456 | return true; | |
3457 | } | |
3458 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
3459 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
3460 | ||
3461 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
3462 | if (slp) | |
3463 | { | |
3464 | strided_load = false; | |
3465 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
a70d6342 IR |
3466 | if (SLP_INSTANCE_LOAD_PERMUTATION (slp_node_instance)) |
3467 | slp_perm = true; | |
3468 | } | |
ebfd146a IR |
3469 | else |
3470 | vec_num = group_size; | |
3471 | ||
3472 | dr_chain = VEC_alloc (tree, heap, vec_num); | |
3473 | } | |
3474 | else | |
3475 | { | |
3476 | first_stmt = stmt; | |
3477 | first_dr = dr; | |
3478 | group_size = vec_num = 1; | |
3479 | } | |
3480 | ||
3481 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr); | |
3482 | gcc_assert (alignment_support_scheme); | |
3483 | ||
3484 | /* In case the vectorization factor (VF) is bigger than the number | |
3485 | of elements that we can fit in a vectype (nunits), we have to generate | |
3486 | more than one vector stmt - i.e - we need to "unroll" the | |
3487 | vector stmt by a factor VF/nunits. In doing so, we record a pointer | |
3488 | from one copy of the vector stmt to the next, in the field | |
3489 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
3490 | stages to find the correct vector defs to be used when vectorizing | |
3491 | stmts that use the defs of the current stmt. The example below illustrates | |
3492 | the vectorization process when VF=16 and nunits=4 (i.e - we need to create | |
3493 | 4 vectorized stmts): | |
3494 | ||
3495 | before vectorization: | |
3496 | RELATED_STMT VEC_STMT | |
3497 | S1: x = memref - - | |
3498 | S2: z = x + 1 - - | |
3499 | ||
3500 | step 1: vectorize stmt S1: | |
3501 | We first create the vector stmt VS1_0, and, as usual, record a | |
3502 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
3503 | Next, we create the vector stmt VS1_1, and record a pointer to | |
3504 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
3505 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of | |
3506 | stmts and pointers: | |
3507 | RELATED_STMT VEC_STMT | |
3508 | VS1_0: vx0 = memref0 VS1_1 - | |
3509 | VS1_1: vx1 = memref1 VS1_2 - | |
3510 | VS1_2: vx2 = memref2 VS1_3 - | |
3511 | VS1_3: vx3 = memref3 - - | |
3512 | S1: x = load - VS1_0 | |
3513 | S2: z = x + 1 - - | |
3514 | ||
b8698a0f L |
3515 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
3516 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
3517 | stmt S2. */ |
3518 | ||
3519 | /* In case of interleaving (non-unit strided access): | |
3520 | ||
3521 | S1: x2 = &base + 2 | |
3522 | S2: x0 = &base | |
3523 | S3: x1 = &base + 1 | |
3524 | S4: x3 = &base + 3 | |
3525 | ||
b8698a0f | 3526 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
3527 | starting from the access of the first stmt of the chain: |
3528 | ||
3529 | VS1: vx0 = &base | |
3530 | VS2: vx1 = &base + vec_size*1 | |
3531 | VS3: vx3 = &base + vec_size*2 | |
3532 | VS4: vx4 = &base + vec_size*3 | |
3533 | ||
3534 | Then permutation statements are generated: | |
3535 | ||
3536 | VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 > | |
3537 | VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 > | |
3538 | ... | |
3539 | ||
3540 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
3541 | (the order of the data-refs in the output of vect_permute_load_chain | |
3542 | corresponds to the order of scalar stmts in the interleaving chain - see | |
3543 | the documentation of vect_permute_load_chain()). | |
3544 | The generation of permutation stmts and recording them in | |
3545 | STMT_VINFO_VEC_STMT is done in vect_transform_strided_load(). | |
3546 | ||
b8698a0f | 3547 | In case of both multiple types and interleaving, the vector loads and |
ebfd146a IR |
3548 | permutation stmts above are created for every copy. The result vector stmts |
3549 | are put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding | |
3550 | STMT_VINFO_RELATED_STMT for the next copies. */ | |
3551 | ||
3552 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
3553 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
3554 | we generate the following code: | |
3555 | p = initial_addr; | |
3556 | indx = 0; | |
3557 | loop { | |
3558 | p = p + indx * vectype_size; | |
3559 | vec_dest = *(p); | |
3560 | indx = indx + 1; | |
3561 | } | |
3562 | ||
3563 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 3564 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
3565 | then generate the following code, in which the data in each iteration is |
3566 | obtained by two vector loads, one from the previous iteration, and one | |
3567 | from the current iteration: | |
3568 | p1 = initial_addr; | |
3569 | msq_init = *(floor(p1)) | |
3570 | p2 = initial_addr + VS - 1; | |
3571 | realignment_token = call target_builtin; | |
3572 | indx = 0; | |
3573 | loop { | |
3574 | p2 = p2 + indx * vectype_size | |
3575 | lsq = *(floor(p2)) | |
3576 | vec_dest = realign_load (msq, lsq, realignment_token) | |
3577 | indx = indx + 1; | |
3578 | msq = lsq; | |
3579 | } */ | |
3580 | ||
3581 | /* If the misalignment remains the same throughout the execution of the | |
3582 | loop, we can create the init_addr and permutation mask at the loop | |
3583 | preheader. Otherwise, it needs to be created inside the loop. | |
3584 | This can only occur when vectorizing memory accesses in the inner-loop | |
3585 | nested within an outer-loop that is being vectorized. */ | |
3586 | ||
a70d6342 | 3587 | if (loop && nested_in_vect_loop_p (loop, stmt) |
ebfd146a IR |
3588 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
3589 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) | |
3590 | { | |
3591 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
3592 | compute_in_loop = true; | |
3593 | } | |
3594 | ||
3595 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
3596 | || alignment_support_scheme == dr_explicit_realign) | |
3597 | && !compute_in_loop) | |
3598 | { | |
3599 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
3600 | alignment_support_scheme, NULL_TREE, | |
3601 | &at_loop); | |
3602 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
3603 | { | |
3604 | phi = SSA_NAME_DEF_STMT (msq); | |
3605 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
3606 | } | |
3607 | } | |
3608 | else | |
3609 | at_loop = loop; | |
3610 | ||
3611 | prev_stmt_info = NULL; | |
3612 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 3613 | { |
ebfd146a IR |
3614 | /* 1. Create the vector pointer update chain. */ |
3615 | if (j == 0) | |
3616 | dataref_ptr = vect_create_data_ref_ptr (first_stmt, | |
b8698a0f L |
3617 | at_loop, offset, |
3618 | &dummy, &ptr_incr, false, | |
5006671f | 3619 | &inv_p); |
ebfd146a | 3620 | else |
b8698a0f | 3621 | dataref_ptr = |
ebfd146a IR |
3622 | bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, NULL_TREE); |
3623 | ||
3624 | for (i = 0; i < vec_num; i++) | |
3625 | { | |
3626 | if (i > 0) | |
3627 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
3628 | NULL_TREE); | |
3629 | ||
3630 | /* 2. Create the vector-load in the loop. */ | |
3631 | switch (alignment_support_scheme) | |
3632 | { | |
3633 | case dr_aligned: | |
3634 | gcc_assert (aligned_access_p (first_dr)); | |
3635 | data_ref = build_fold_indirect_ref (dataref_ptr); | |
3636 | break; | |
3637 | case dr_unaligned_supported: | |
3638 | { | |
3639 | int mis = DR_MISALIGNMENT (first_dr); | |
3640 | tree tmis = (mis == -1 ? size_zero_node : size_int (mis)); | |
3641 | ||
3642 | tmis = size_binop (MULT_EXPR, tmis, size_int(BITS_PER_UNIT)); | |
3643 | data_ref = | |
3644 | build2 (MISALIGNED_INDIRECT_REF, vectype, dataref_ptr, tmis); | |
3645 | break; | |
3646 | } | |
3647 | case dr_explicit_realign: | |
3648 | { | |
3649 | tree ptr, bump; | |
3650 | tree vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
3651 | ||
3652 | if (compute_in_loop) | |
3653 | msq = vect_setup_realignment (first_stmt, gsi, | |
3654 | &realignment_token, | |
b8698a0f | 3655 | dr_explicit_realign, |
ebfd146a IR |
3656 | dataref_ptr, NULL); |
3657 | ||
3658 | data_ref = build1 (ALIGN_INDIRECT_REF, vectype, dataref_ptr); | |
3659 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3660 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
3661 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3662 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5006671f RG |
3663 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); |
3664 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
ebfd146a | 3665 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
3666 | msq = new_temp; |
3667 | ||
3668 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
3669 | TYPE_SIZE_UNIT (scalar_type)); | |
3670 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); | |
3671 | data_ref = build1 (ALIGN_INDIRECT_REF, vectype, ptr); | |
3672 | break; | |
3673 | } | |
3674 | case dr_explicit_realign_optimized: | |
3675 | data_ref = build1 (ALIGN_INDIRECT_REF, vectype, dataref_ptr); | |
3676 | break; | |
3677 | default: | |
3678 | gcc_unreachable (); | |
3679 | } | |
5006671f RG |
3680 | /* If accesses through a pointer to vectype do not alias the original |
3681 | memory reference we have a problem. This should never happen. */ | |
3682 | gcc_assert (alias_sets_conflict_p (get_alias_set (data_ref), | |
3683 | get_alias_set (gimple_assign_rhs1 (stmt)))); | |
ebfd146a IR |
3684 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
3685 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
3686 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3687 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3688 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3689 | mark_symbols_for_renaming (new_stmt); | |
3690 | ||
3691 | /* 3. Handle explicit realignment if necessary/supported. Create in | |
3692 | loop: vec_dest = realign_load (msq, lsq, realignment_token) */ | |
3693 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
3694 | || alignment_support_scheme == dr_explicit_realign) | |
3695 | { | |
3696 | tree tmp; | |
3697 | ||
3698 | lsq = gimple_assign_lhs (new_stmt); | |
3699 | if (!realignment_token) | |
3700 | realignment_token = dataref_ptr; | |
3701 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3702 | tmp = build3 (REALIGN_LOAD_EXPR, vectype, msq, lsq, | |
3703 | realignment_token); | |
3704 | new_stmt = gimple_build_assign (vec_dest, tmp); | |
3705 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3706 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3707 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3708 | ||
3709 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
3710 | { | |
3711 | gcc_assert (phi); | |
3712 | if (i == vec_num - 1 && j == ncopies - 1) | |
f5045c96 AM |
3713 | add_phi_arg (phi, lsq, loop_latch_edge (containing_loop), |
3714 | UNKNOWN_LOCATION); | |
ebfd146a IR |
3715 | msq = lsq; |
3716 | } | |
3717 | } | |
3718 | ||
3719 | /* 4. Handle invariant-load. */ | |
a70d6342 | 3720 | if (inv_p && !bb_vinfo) |
ebfd146a IR |
3721 | { |
3722 | gcc_assert (!strided_load); | |
3723 | gcc_assert (nested_in_vect_loop_p (loop, stmt)); | |
3724 | if (j == 0) | |
3725 | { | |
3726 | int k; | |
3727 | tree t = NULL_TREE; | |
3728 | tree vec_inv, bitpos, bitsize = TYPE_SIZE (scalar_type); | |
3729 | ||
3730 | /* CHECKME: bitpos depends on endianess? */ | |
3731 | bitpos = bitsize_zero_node; | |
b8698a0f | 3732 | vec_inv = build3 (BIT_FIELD_REF, scalar_type, new_temp, |
ebfd146a | 3733 | bitsize, bitpos); |
b8698a0f | 3734 | vec_dest = |
ebfd146a IR |
3735 | vect_create_destination_var (scalar_dest, NULL_TREE); |
3736 | new_stmt = gimple_build_assign (vec_dest, vec_inv); | |
3737 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3738 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3739 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3740 | ||
3741 | for (k = nunits - 1; k >= 0; --k) | |
3742 | t = tree_cons (NULL_TREE, new_temp, t); | |
3743 | /* FIXME: use build_constructor directly. */ | |
3744 | vec_inv = build_constructor_from_list (vectype, t); | |
3745 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
3746 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
3747 | } | |
3748 | else | |
3749 | gcc_unreachable (); /* FORNOW. */ | |
3750 | } | |
3751 | ||
3752 | /* Collect vector loads and later create their permutation in | |
3753 | vect_transform_strided_load (). */ | |
3754 | if (strided_load || slp_perm) | |
3755 | VEC_quick_push (tree, dr_chain, new_temp); | |
3756 | ||
3757 | /* Store vector loads in the corresponding SLP_NODE. */ | |
3758 | if (slp && !slp_perm) | |
3759 | VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); | |
3760 | } | |
3761 | ||
3762 | if (slp && !slp_perm) | |
3763 | continue; | |
3764 | ||
3765 | if (slp_perm) | |
3766 | { | |
a70d6342 | 3767 | if (!vect_transform_slp_perm_load (stmt, dr_chain, gsi, vf, |
ebfd146a IR |
3768 | slp_node_instance, false)) |
3769 | { | |
3770 | VEC_free (tree, heap, dr_chain); | |
3771 | return false; | |
3772 | } | |
3773 | } | |
3774 | else | |
3775 | { | |
3776 | if (strided_load) | |
3777 | { | |
3778 | if (!vect_transform_strided_load (stmt, dr_chain, group_size, gsi)) | |
b8698a0f | 3779 | return false; |
ebfd146a IR |
3780 | |
3781 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
3782 | VEC_free (tree, heap, dr_chain); | |
3783 | dr_chain = VEC_alloc (tree, heap, group_size); | |
3784 | } | |
3785 | else | |
3786 | { | |
3787 | if (j == 0) | |
3788 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3789 | else | |
3790 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3791 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3792 | } | |
3793 | } | |
3794 | } | |
3795 | ||
3796 | if (dr_chain) | |
3797 | VEC_free (tree, heap, dr_chain); | |
3798 | ||
3799 | return true; | |
3800 | } | |
3801 | ||
3802 | /* Function vect_is_simple_cond. | |
b8698a0f | 3803 | |
ebfd146a IR |
3804 | Input: |
3805 | LOOP - the loop that is being vectorized. | |
3806 | COND - Condition that is checked for simple use. | |
3807 | ||
3808 | Returns whether a COND can be vectorized. Checks whether | |
3809 | condition operands are supportable using vec_is_simple_use. */ | |
3810 | ||
3811 | static bool | |
3812 | vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo) | |
3813 | { | |
3814 | tree lhs, rhs; | |
3815 | tree def; | |
3816 | enum vect_def_type dt; | |
3817 | ||
3818 | if (!COMPARISON_CLASS_P (cond)) | |
3819 | return false; | |
3820 | ||
3821 | lhs = TREE_OPERAND (cond, 0); | |
3822 | rhs = TREE_OPERAND (cond, 1); | |
3823 | ||
3824 | if (TREE_CODE (lhs) == SSA_NAME) | |
3825 | { | |
3826 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
b8698a0f | 3827 | if (!vect_is_simple_use (lhs, loop_vinfo, NULL, &lhs_def_stmt, &def, |
a70d6342 | 3828 | &dt)) |
ebfd146a IR |
3829 | return false; |
3830 | } | |
3831 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
3832 | && TREE_CODE (lhs) != FIXED_CST) | |
3833 | return false; | |
3834 | ||
3835 | if (TREE_CODE (rhs) == SSA_NAME) | |
3836 | { | |
3837 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
b8698a0f | 3838 | if (!vect_is_simple_use (rhs, loop_vinfo, NULL, &rhs_def_stmt, &def, |
a70d6342 | 3839 | &dt)) |
ebfd146a IR |
3840 | return false; |
3841 | } | |
3842 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST | |
3843 | && TREE_CODE (rhs) != FIXED_CST) | |
3844 | return false; | |
3845 | ||
3846 | return true; | |
3847 | } | |
3848 | ||
3849 | /* vectorizable_condition. | |
3850 | ||
b8698a0f L |
3851 | Check if STMT is conditional modify expression that can be vectorized. |
3852 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3853 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
3854 | at GSI. |
3855 | ||
3856 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
3857 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
3858 | else caluse if it is 2). | |
ebfd146a IR |
3859 | |
3860 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3861 | ||
4bbe8262 | 3862 | bool |
ebfd146a | 3863 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
4bbe8262 | 3864 | gimple *vec_stmt, tree reduc_def, int reduc_index) |
ebfd146a IR |
3865 | { |
3866 | tree scalar_dest = NULL_TREE; | |
3867 | tree vec_dest = NULL_TREE; | |
3868 | tree op = NULL_TREE; | |
3869 | tree cond_expr, then_clause, else_clause; | |
3870 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3871 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3872 | tree vec_cond_lhs, vec_cond_rhs, vec_then_clause, vec_else_clause; | |
3873 | tree vec_compare, vec_cond_expr; | |
3874 | tree new_temp; | |
3875 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3876 | enum machine_mode vec_mode; | |
3877 | tree def; | |
3878 | enum vect_def_type dt; | |
3879 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
3880 | int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3881 | enum tree_code code; | |
3882 | ||
a70d6342 IR |
3883 | /* FORNOW: unsupported in basic block SLP. */ |
3884 | gcc_assert (loop_vinfo); | |
b8698a0f | 3885 | |
ebfd146a IR |
3886 | gcc_assert (ncopies >= 1); |
3887 | if (ncopies > 1) | |
3888 | return false; /* FORNOW */ | |
3889 | ||
3890 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
3891 | return false; | |
3892 | ||
4bbe8262 IR |
3893 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3894 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
3895 | && reduc_def)) | |
ebfd146a IR |
3896 | return false; |
3897 | ||
3898 | /* FORNOW: SLP not supported. */ | |
3899 | if (STMT_SLP_TYPE (stmt_info)) | |
3900 | return false; | |
3901 | ||
3902 | /* FORNOW: not yet supported. */ | |
b8698a0f | 3903 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a IR |
3904 | { |
3905 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3906 | fprintf (vect_dump, "value used after loop."); | |
3907 | return false; | |
3908 | } | |
3909 | ||
3910 | /* Is vectorizable conditional operation? */ | |
3911 | if (!is_gimple_assign (stmt)) | |
3912 | return false; | |
3913 | ||
3914 | code = gimple_assign_rhs_code (stmt); | |
3915 | ||
3916 | if (code != COND_EXPR) | |
3917 | return false; | |
3918 | ||
3919 | gcc_assert (gimple_assign_single_p (stmt)); | |
3920 | op = gimple_assign_rhs1 (stmt); | |
3921 | cond_expr = TREE_OPERAND (op, 0); | |
3922 | then_clause = TREE_OPERAND (op, 1); | |
3923 | else_clause = TREE_OPERAND (op, 2); | |
3924 | ||
3925 | if (!vect_is_simple_cond (cond_expr, loop_vinfo)) | |
3926 | return false; | |
3927 | ||
3928 | /* We do not handle two different vector types for the condition | |
3929 | and the values. */ | |
8533c9d8 SP |
3930 | if (!types_compatible_p (TREE_TYPE (TREE_OPERAND (cond_expr, 0)), |
3931 | TREE_TYPE (vectype))) | |
ebfd146a IR |
3932 | return false; |
3933 | ||
3934 | if (TREE_CODE (then_clause) == SSA_NAME) | |
3935 | { | |
3936 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
b8698a0f | 3937 | if (!vect_is_simple_use (then_clause, loop_vinfo, NULL, |
ebfd146a IR |
3938 | &then_def_stmt, &def, &dt)) |
3939 | return false; | |
3940 | } | |
b8698a0f | 3941 | else if (TREE_CODE (then_clause) != INTEGER_CST |
ebfd146a IR |
3942 | && TREE_CODE (then_clause) != REAL_CST |
3943 | && TREE_CODE (then_clause) != FIXED_CST) | |
3944 | return false; | |
3945 | ||
3946 | if (TREE_CODE (else_clause) == SSA_NAME) | |
3947 | { | |
3948 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
a70d6342 | 3949 | if (!vect_is_simple_use (else_clause, loop_vinfo, NULL, |
ebfd146a IR |
3950 | &else_def_stmt, &def, &dt)) |
3951 | return false; | |
3952 | } | |
b8698a0f | 3953 | else if (TREE_CODE (else_clause) != INTEGER_CST |
ebfd146a IR |
3954 | && TREE_CODE (else_clause) != REAL_CST |
3955 | && TREE_CODE (else_clause) != FIXED_CST) | |
3956 | return false; | |
3957 | ||
3958 | ||
3959 | vec_mode = TYPE_MODE (vectype); | |
3960 | ||
b8698a0f | 3961 | if (!vec_stmt) |
ebfd146a IR |
3962 | { |
3963 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
8e7aa1f9 | 3964 | return expand_vec_cond_expr_p (TREE_TYPE (op), vec_mode); |
ebfd146a IR |
3965 | } |
3966 | ||
3967 | /* Transform */ | |
3968 | ||
3969 | /* Handle def. */ | |
3970 | scalar_dest = gimple_assign_lhs (stmt); | |
3971 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3972 | ||
3973 | /* Handle cond expr. */ | |
b8698a0f | 3974 | vec_cond_lhs = |
ebfd146a | 3975 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), stmt, NULL); |
b8698a0f | 3976 | vec_cond_rhs = |
ebfd146a | 3977 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), stmt, NULL); |
4bbe8262 IR |
3978 | if (reduc_index == 1) |
3979 | vec_then_clause = reduc_def; | |
3980 | else | |
3981 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, stmt, NULL); | |
3982 | if (reduc_index == 2) | |
3983 | vec_else_clause = reduc_def; | |
3984 | else | |
3985 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, stmt, NULL); | |
ebfd146a IR |
3986 | |
3987 | /* Arguments are ready. Create the new vector stmt. */ | |
b8698a0f | 3988 | vec_compare = build2 (TREE_CODE (cond_expr), vectype, |
ebfd146a | 3989 | vec_cond_lhs, vec_cond_rhs); |
b8698a0f | 3990 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, |
ebfd146a IR |
3991 | vec_compare, vec_then_clause, vec_else_clause); |
3992 | ||
3993 | *vec_stmt = gimple_build_assign (vec_dest, vec_cond_expr); | |
3994 | new_temp = make_ssa_name (vec_dest, *vec_stmt); | |
3995 | gimple_assign_set_lhs (*vec_stmt, new_temp); | |
3996 | vect_finish_stmt_generation (stmt, *vec_stmt, gsi); | |
b8698a0f | 3997 | |
ebfd146a IR |
3998 | return true; |
3999 | } | |
4000 | ||
4001 | ||
8644a673 | 4002 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
4003 | |
4004 | bool | |
a70d6342 | 4005 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 4006 | { |
8644a673 | 4007 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 4008 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 4009 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 4010 | bool ok; |
a70d6342 | 4011 | tree scalar_type, vectype; |
ebfd146a IR |
4012 | |
4013 | if (vect_print_dump_info (REPORT_DETAILS)) | |
ebfd146a | 4014 | { |
8644a673 IR |
4015 | fprintf (vect_dump, "==> examining statement: "); |
4016 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4017 | } | |
ebfd146a | 4018 | |
1825a1f3 | 4019 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 4020 | { |
1825a1f3 IR |
4021 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4022 | fprintf (vect_dump, "not vectorized: stmt has volatile operands"); | |
4023 | ||
4024 | return false; | |
4025 | } | |
b8698a0f L |
4026 | |
4027 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
4028 | to include: |
4029 | - the COND_EXPR which is the loop exit condition | |
4030 | - any LABEL_EXPRs in the loop | |
b8698a0f | 4031 | - computations that are used only for array indexing or loop control. |
8644a673 IR |
4032 | In basic blocks we only analyze statements that are a part of some SLP |
4033 | instance, therefore, all the statements are relevant. */ | |
ebfd146a | 4034 | |
b8698a0f | 4035 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 4036 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a IR |
4037 | { |
4038 | if (vect_print_dump_info (REPORT_DETAILS)) | |
8644a673 | 4039 | fprintf (vect_dump, "irrelevant."); |
ebfd146a | 4040 | |
8644a673 IR |
4041 | return true; |
4042 | } | |
ebfd146a | 4043 | |
8644a673 IR |
4044 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
4045 | { | |
4046 | case vect_internal_def: | |
4047 | break; | |
ebfd146a | 4048 | |
8644a673 | 4049 | case vect_reduction_def: |
7c5222ff | 4050 | case vect_nested_cycle: |
a70d6342 | 4051 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 4052 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 4053 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
4054 | break; |
4055 | ||
4056 | case vect_induction_def: | |
4057 | case vect_constant_def: | |
4058 | case vect_external_def: | |
4059 | case vect_unknown_def_type: | |
4060 | default: | |
4061 | gcc_unreachable (); | |
4062 | } | |
ebfd146a | 4063 | |
a70d6342 IR |
4064 | if (bb_vinfo) |
4065 | { | |
4066 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
4067 | ||
b690cc0f | 4068 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
a70d6342 IR |
4069 | if (vect_print_dump_info (REPORT_DETAILS)) |
4070 | { | |
4071 | fprintf (vect_dump, "get vectype for scalar type: "); | |
4072 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4073 | } | |
4074 | ||
4075 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4076 | if (!vectype) | |
4077 | { | |
4078 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4079 | { | |
4080 | fprintf (vect_dump, "not SLPed: unsupported data-type "); | |
4081 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4082 | } | |
4083 | return false; | |
4084 | } | |
4085 | ||
4086 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4087 | { | |
4088 | fprintf (vect_dump, "vectype: "); | |
4089 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
4090 | } | |
4091 | ||
4092 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
4093 | } | |
4094 | ||
8644a673 | 4095 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 4096 | { |
8644a673 IR |
4097 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
4098 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); | |
4099 | *need_to_vectorize = true; | |
ebfd146a IR |
4100 | } |
4101 | ||
8644a673 | 4102 | ok = true; |
b8698a0f | 4103 | if (!bb_vinfo |
a70d6342 IR |
4104 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
4105 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8644a673 IR |
4106 | ok = (vectorizable_type_promotion (stmt, NULL, NULL, NULL) |
4107 | || vectorizable_type_demotion (stmt, NULL, NULL, NULL) | |
4108 | || vectorizable_conversion (stmt, NULL, NULL, NULL) | |
4109 | || vectorizable_operation (stmt, NULL, NULL, NULL) | |
4110 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
4111 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
4112 | || vectorizable_call (stmt, NULL, NULL) | |
4113 | || vectorizable_store (stmt, NULL, NULL, NULL) | |
b5aeb3bb | 4114 | || vectorizable_reduction (stmt, NULL, NULL, NULL) |
4bbe8262 | 4115 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0)); |
a70d6342 IR |
4116 | else |
4117 | { | |
4118 | if (bb_vinfo) | |
4119 | ok = (vectorizable_operation (stmt, NULL, NULL, node) | |
4120 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
4121 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
4122 | || vectorizable_store (stmt, NULL, NULL, node)); | |
b8698a0f | 4123 | } |
8644a673 IR |
4124 | |
4125 | if (!ok) | |
ebfd146a | 4126 | { |
8644a673 IR |
4127 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4128 | { | |
4129 | fprintf (vect_dump, "not vectorized: relevant stmt not "); | |
4130 | fprintf (vect_dump, "supported: "); | |
4131 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4132 | } | |
b8698a0f | 4133 | |
ebfd146a IR |
4134 | return false; |
4135 | } | |
4136 | ||
a70d6342 IR |
4137 | if (bb_vinfo) |
4138 | return true; | |
4139 | ||
8644a673 IR |
4140 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
4141 | need extra handling, except for vectorizable reductions. */ | |
4142 | if (STMT_VINFO_LIVE_P (stmt_info) | |
4143 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
4144 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 4145 | |
8644a673 | 4146 | if (!ok) |
ebfd146a | 4147 | { |
8644a673 IR |
4148 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
4149 | { | |
4150 | fprintf (vect_dump, "not vectorized: live stmt not "); | |
4151 | fprintf (vect_dump, "supported: "); | |
4152 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4153 | } | |
b8698a0f | 4154 | |
8644a673 | 4155 | return false; |
ebfd146a IR |
4156 | } |
4157 | ||
8644a673 | 4158 | if (!PURE_SLP_STMT (stmt_info)) |
ebfd146a | 4159 | { |
b8698a0f L |
4160 | /* Groups of strided accesses whose size is not a power of 2 are not |
4161 | vectorizable yet using loop-vectorization. Therefore, if this stmt | |
4162 | feeds non-SLP-able stmts (i.e., this stmt has to be both SLPed and | |
a70d6342 | 4163 | loop-based vectorized), the loop cannot be vectorized. */ |
8644a673 IR |
4164 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) |
4165 | && exact_log2 (DR_GROUP_SIZE (vinfo_for_stmt ( | |
4166 | DR_GROUP_FIRST_DR (stmt_info)))) == -1) | |
ebfd146a | 4167 | { |
8644a673 IR |
4168 | if (vect_print_dump_info (REPORT_DETAILS)) |
4169 | { | |
4170 | fprintf (vect_dump, "not vectorized: the size of group " | |
4171 | "of strided accesses is not a power of 2"); | |
4172 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
4173 | } | |
4174 | ||
ebfd146a IR |
4175 | return false; |
4176 | } | |
4177 | } | |
b8698a0f | 4178 | |
ebfd146a IR |
4179 | return true; |
4180 | } | |
4181 | ||
4182 | ||
4183 | /* Function vect_transform_stmt. | |
4184 | ||
4185 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
4186 | ||
4187 | bool | |
4188 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
b8698a0f | 4189 | bool *strided_store, slp_tree slp_node, |
ebfd146a IR |
4190 | slp_instance slp_node_instance) |
4191 | { | |
4192 | bool is_store = false; | |
4193 | gimple vec_stmt = NULL; | |
4194 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4195 | gimple orig_stmt_in_pattern; | |
4196 | bool done; | |
ebfd146a IR |
4197 | |
4198 | switch (STMT_VINFO_TYPE (stmt_info)) | |
4199 | { | |
4200 | case type_demotion_vec_info_type: | |
4201 | done = vectorizable_type_demotion (stmt, gsi, &vec_stmt, slp_node); | |
4202 | gcc_assert (done); | |
4203 | break; | |
4204 | ||
4205 | case type_promotion_vec_info_type: | |
4206 | done = vectorizable_type_promotion (stmt, gsi, &vec_stmt, slp_node); | |
4207 | gcc_assert (done); | |
4208 | break; | |
4209 | ||
4210 | case type_conversion_vec_info_type: | |
4211 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
4212 | gcc_assert (done); | |
4213 | break; | |
4214 | ||
4215 | case induc_vec_info_type: | |
4216 | gcc_assert (!slp_node); | |
4217 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
4218 | gcc_assert (done); | |
4219 | break; | |
4220 | ||
4221 | case op_vec_info_type: | |
4222 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
4223 | gcc_assert (done); | |
4224 | break; | |
4225 | ||
4226 | case assignment_vec_info_type: | |
4227 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
4228 | gcc_assert (done); | |
4229 | break; | |
4230 | ||
4231 | case load_vec_info_type: | |
b8698a0f | 4232 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
4233 | slp_node_instance); |
4234 | gcc_assert (done); | |
4235 | break; | |
4236 | ||
4237 | case store_vec_info_type: | |
4238 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
4239 | gcc_assert (done); | |
4240 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) && !slp_node) | |
4241 | { | |
4242 | /* In case of interleaving, the whole chain is vectorized when the | |
4243 | last store in the chain is reached. Store stmts before the last | |
4244 | one are skipped, and there vec_stmt_info shouldn't be freed | |
4245 | meanwhile. */ | |
4246 | *strided_store = true; | |
4247 | if (STMT_VINFO_VEC_STMT (stmt_info)) | |
4248 | is_store = true; | |
4249 | } | |
4250 | else | |
4251 | is_store = true; | |
4252 | break; | |
4253 | ||
4254 | case condition_vec_info_type: | |
4255 | gcc_assert (!slp_node); | |
4bbe8262 | 4256 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0); |
ebfd146a IR |
4257 | gcc_assert (done); |
4258 | break; | |
4259 | ||
4260 | case call_vec_info_type: | |
4261 | gcc_assert (!slp_node); | |
4262 | done = vectorizable_call (stmt, gsi, &vec_stmt); | |
4263 | break; | |
4264 | ||
4265 | case reduc_vec_info_type: | |
b5aeb3bb | 4266 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
4267 | gcc_assert (done); |
4268 | break; | |
4269 | ||
4270 | default: | |
4271 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
4272 | { | |
4273 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4274 | fprintf (vect_dump, "stmt not supported."); | |
4275 | gcc_unreachable (); | |
4276 | } | |
4277 | } | |
4278 | ||
4279 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
4280 | is being vectorized, but outside the immediately enclosing loop. */ | |
4281 | if (vec_stmt | |
a70d6342 IR |
4282 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
4283 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
4284 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
4285 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
4286 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 4287 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 4288 | vect_used_in_outer_by_reduction)) |
ebfd146a | 4289 | { |
a70d6342 IR |
4290 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
4291 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
4292 | imm_use_iterator imm_iter; |
4293 | use_operand_p use_p; | |
4294 | tree scalar_dest; | |
4295 | gimple exit_phi; | |
4296 | ||
4297 | if (vect_print_dump_info (REPORT_DETAILS)) | |
a70d6342 | 4298 | fprintf (vect_dump, "Record the vdef for outer-loop vectorization."); |
ebfd146a IR |
4299 | |
4300 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
4301 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
4302 | STMT). */ | |
4303 | if (gimple_code (stmt) == GIMPLE_PHI) | |
4304 | scalar_dest = PHI_RESULT (stmt); | |
4305 | else | |
4306 | scalar_dest = gimple_assign_lhs (stmt); | |
4307 | ||
4308 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
4309 | { | |
4310 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
4311 | { | |
4312 | exit_phi = USE_STMT (use_p); | |
4313 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
4314 | } | |
4315 | } | |
4316 | } | |
4317 | ||
4318 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
4319 | being vectorized. */ | |
4320 | if (STMT_VINFO_LIVE_P (stmt_info) | |
4321 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
4322 | { | |
4323 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
4324 | gcc_assert (done); | |
4325 | } | |
4326 | ||
4327 | if (vec_stmt) | |
4328 | { | |
4329 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; | |
4330 | orig_stmt_in_pattern = STMT_VINFO_RELATED_STMT (stmt_info); | |
4331 | if (orig_stmt_in_pattern) | |
4332 | { | |
4333 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt_in_pattern); | |
4334 | /* STMT was inserted by the vectorizer to replace a computation idiom. | |
b8698a0f L |
4335 | ORIG_STMT_IN_PATTERN is a stmt in the original sequence that |
4336 | computed this idiom. We need to record a pointer to VEC_STMT in | |
4337 | the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the | |
ebfd146a IR |
4338 | documentation of vect_pattern_recog. */ |
4339 | if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)) | |
4340 | { | |
4341 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt); | |
4342 | STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt; | |
4343 | } | |
4344 | } | |
4345 | } | |
4346 | ||
b8698a0f | 4347 | return is_store; |
ebfd146a IR |
4348 | } |
4349 | ||
4350 | ||
b8698a0f | 4351 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
4352 | stmt_vec_info. */ |
4353 | ||
4354 | void | |
4355 | vect_remove_stores (gimple first_stmt) | |
4356 | { | |
4357 | gimple next = first_stmt; | |
4358 | gimple tmp; | |
4359 | gimple_stmt_iterator next_si; | |
4360 | ||
4361 | while (next) | |
4362 | { | |
4363 | /* Free the attached stmt_vec_info and remove the stmt. */ | |
4364 | next_si = gsi_for_stmt (next); | |
4365 | gsi_remove (&next_si, true); | |
4366 | tmp = DR_GROUP_NEXT_DR (vinfo_for_stmt (next)); | |
4367 | free_stmt_vec_info (next); | |
4368 | next = tmp; | |
4369 | } | |
4370 | } | |
4371 | ||
4372 | ||
4373 | /* Function new_stmt_vec_info. | |
4374 | ||
4375 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
4376 | ||
4377 | stmt_vec_info | |
b8698a0f | 4378 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 4379 | bb_vec_info bb_vinfo) |
ebfd146a IR |
4380 | { |
4381 | stmt_vec_info res; | |
4382 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
4383 | ||
4384 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
4385 | STMT_VINFO_STMT (res) = stmt; | |
4386 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 4387 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 4388 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
4389 | STMT_VINFO_LIVE_P (res) = false; |
4390 | STMT_VINFO_VECTYPE (res) = NULL; | |
4391 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 4392 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
4393 | STMT_VINFO_IN_PATTERN_P (res) = false; |
4394 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
4395 | STMT_VINFO_DATA_REF (res) = NULL; | |
4396 | ||
4397 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
4398 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
4399 | STMT_VINFO_DR_INIT (res) = NULL; | |
4400 | STMT_VINFO_DR_STEP (res) = NULL; | |
4401 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
4402 | ||
4403 | if (gimple_code (stmt) == GIMPLE_PHI | |
4404 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
4405 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
4406 | else | |
8644a673 IR |
4407 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
4408 | ||
ebfd146a IR |
4409 | STMT_VINFO_SAME_ALIGN_REFS (res) = VEC_alloc (dr_p, heap, 5); |
4410 | STMT_VINFO_INSIDE_OF_LOOP_COST (res) = 0; | |
4411 | STMT_VINFO_OUTSIDE_OF_LOOP_COST (res) = 0; | |
32e8bb8e | 4412 | STMT_SLP_TYPE (res) = loop_vect; |
ebfd146a IR |
4413 | DR_GROUP_FIRST_DR (res) = NULL; |
4414 | DR_GROUP_NEXT_DR (res) = NULL; | |
4415 | DR_GROUP_SIZE (res) = 0; | |
4416 | DR_GROUP_STORE_COUNT (res) = 0; | |
4417 | DR_GROUP_GAP (res) = 0; | |
4418 | DR_GROUP_SAME_DR_STMT (res) = NULL; | |
4419 | DR_GROUP_READ_WRITE_DEPENDENCE (res) = false; | |
4420 | ||
4421 | return res; | |
4422 | } | |
4423 | ||
4424 | ||
4425 | /* Create a hash table for stmt_vec_info. */ | |
4426 | ||
4427 | void | |
4428 | init_stmt_vec_info_vec (void) | |
4429 | { | |
4430 | gcc_assert (!stmt_vec_info_vec); | |
4431 | stmt_vec_info_vec = VEC_alloc (vec_void_p, heap, 50); | |
4432 | } | |
4433 | ||
4434 | ||
4435 | /* Free hash table for stmt_vec_info. */ | |
4436 | ||
4437 | void | |
4438 | free_stmt_vec_info_vec (void) | |
4439 | { | |
4440 | gcc_assert (stmt_vec_info_vec); | |
4441 | VEC_free (vec_void_p, heap, stmt_vec_info_vec); | |
4442 | } | |
4443 | ||
4444 | ||
4445 | /* Free stmt vectorization related info. */ | |
4446 | ||
4447 | void | |
4448 | free_stmt_vec_info (gimple stmt) | |
4449 | { | |
4450 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4451 | ||
4452 | if (!stmt_info) | |
4453 | return; | |
4454 | ||
4455 | VEC_free (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmt_info)); | |
4456 | set_vinfo_for_stmt (stmt, NULL); | |
4457 | free (stmt_info); | |
4458 | } | |
4459 | ||
4460 | ||
4461 | /* Function get_vectype_for_scalar_type. | |
4462 | ||
4463 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
4464 | by the target. */ | |
4465 | ||
4466 | tree | |
4467 | get_vectype_for_scalar_type (tree scalar_type) | |
4468 | { | |
4469 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
2f816591 | 4470 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
4471 | int nunits; |
4472 | tree vectype; | |
4473 | ||
4474 | if (nbytes == 0 || nbytes >= UNITS_PER_SIMD_WORD (inner_mode)) | |
4475 | return NULL_TREE; | |
4476 | ||
2f816591 RG |
4477 | /* We can't build a vector type of elements with alignment bigger than |
4478 | their size. */ | |
4479 | if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) | |
4480 | return NULL_TREE; | |
4481 | ||
6d7971b8 RG |
4482 | /* If we'd build a vector type of elements whose mode precision doesn't |
4483 | match their types precision we'll get mismatched types on vector | |
4484 | extracts via BIT_FIELD_REFs. This effectively means we disable | |
4485 | vectorization of bool and/or enum types in some languages. */ | |
4486 | if (INTEGRAL_TYPE_P (scalar_type) | |
4487 | && GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type)) | |
4488 | return NULL_TREE; | |
4489 | ||
ebfd146a IR |
4490 | /* FORNOW: Only a single vector size per mode (UNITS_PER_SIMD_WORD) |
4491 | is expected. */ | |
4492 | nunits = UNITS_PER_SIMD_WORD (inner_mode) / nbytes; | |
4493 | ||
4494 | vectype = build_vector_type (scalar_type, nunits); | |
4495 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4496 | { | |
4497 | fprintf (vect_dump, "get vectype with %d units of type ", nunits); | |
4498 | print_generic_expr (vect_dump, scalar_type, TDF_SLIM); | |
4499 | } | |
4500 | ||
4501 | if (!vectype) | |
4502 | return NULL_TREE; | |
4503 | ||
4504 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4505 | { | |
4506 | fprintf (vect_dump, "vectype: "); | |
4507 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
4508 | } | |
4509 | ||
4510 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
4511 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
4512 | { | |
4513 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4514 | fprintf (vect_dump, "mode not supported by target."); | |
4515 | return NULL_TREE; | |
4516 | } | |
4517 | ||
4518 | return vectype; | |
4519 | } | |
4520 | ||
b690cc0f RG |
4521 | /* Function get_same_sized_vectype |
4522 | ||
4523 | Returns a vector type corresponding to SCALAR_TYPE of size | |
4524 | VECTOR_TYPE if supported by the target. */ | |
4525 | ||
4526 | tree | |
4527 | get_same_sized_vectype (tree scalar_type, tree vector_type ATTRIBUTE_UNUSED) | |
4528 | { | |
4529 | return get_vectype_for_scalar_type (scalar_type); | |
4530 | } | |
4531 | ||
ebfd146a IR |
4532 | /* Function vect_is_simple_use. |
4533 | ||
4534 | Input: | |
a70d6342 IR |
4535 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
4536 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
4537 | OPERAND - operand of a stmt in the loop or bb. | |
ebfd146a IR |
4538 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
4539 | ||
4540 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f L |
4541 | For loops, supportable operands are constants, loop invariants, and operands |
4542 | that are defined by the current iteration of the loop. Unsupportable | |
4543 | operands are those that are defined by a previous iteration of the loop (as | |
a70d6342 IR |
4544 | is the case in reduction/induction computations). |
4545 | For basic blocks, supportable operands are constants and bb invariants. | |
4546 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
4547 | |
4548 | bool | |
b8698a0f | 4549 | vect_is_simple_use (tree operand, loop_vec_info loop_vinfo, |
a70d6342 | 4550 | bb_vec_info bb_vinfo, gimple *def_stmt, |
ebfd146a | 4551 | tree *def, enum vect_def_type *dt) |
b8698a0f | 4552 | { |
ebfd146a IR |
4553 | basic_block bb; |
4554 | stmt_vec_info stmt_vinfo; | |
a70d6342 | 4555 | struct loop *loop = NULL; |
b8698a0f | 4556 | |
a70d6342 IR |
4557 | if (loop_vinfo) |
4558 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
4559 | |
4560 | *def_stmt = NULL; | |
4561 | *def = NULL_TREE; | |
b8698a0f | 4562 | |
ebfd146a IR |
4563 | if (vect_print_dump_info (REPORT_DETAILS)) |
4564 | { | |
4565 | fprintf (vect_dump, "vect_is_simple_use: operand "); | |
4566 | print_generic_expr (vect_dump, operand, TDF_SLIM); | |
4567 | } | |
b8698a0f | 4568 | |
ebfd146a IR |
4569 | if (TREE_CODE (operand) == INTEGER_CST || TREE_CODE (operand) == REAL_CST) |
4570 | { | |
4571 | *dt = vect_constant_def; | |
4572 | return true; | |
4573 | } | |
b8698a0f | 4574 | |
ebfd146a IR |
4575 | if (is_gimple_min_invariant (operand)) |
4576 | { | |
4577 | *def = operand; | |
8644a673 | 4578 | *dt = vect_external_def; |
ebfd146a IR |
4579 | return true; |
4580 | } | |
4581 | ||
4582 | if (TREE_CODE (operand) == PAREN_EXPR) | |
4583 | { | |
4584 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4585 | fprintf (vect_dump, "non-associatable copy."); | |
4586 | operand = TREE_OPERAND (operand, 0); | |
4587 | } | |
b8698a0f | 4588 | |
ebfd146a IR |
4589 | if (TREE_CODE (operand) != SSA_NAME) |
4590 | { | |
4591 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4592 | fprintf (vect_dump, "not ssa-name."); | |
4593 | return false; | |
4594 | } | |
b8698a0f | 4595 | |
ebfd146a IR |
4596 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
4597 | if (*def_stmt == NULL) | |
4598 | { | |
4599 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4600 | fprintf (vect_dump, "no def_stmt."); | |
4601 | return false; | |
4602 | } | |
4603 | ||
4604 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4605 | { | |
4606 | fprintf (vect_dump, "def_stmt: "); | |
4607 | print_gimple_stmt (vect_dump, *def_stmt, 0, TDF_SLIM); | |
4608 | } | |
4609 | ||
8644a673 | 4610 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
4611 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
4612 | if (gimple_nop_p (*def_stmt)) | |
4613 | { | |
4614 | *def = operand; | |
8644a673 | 4615 | *dt = vect_external_def; |
ebfd146a IR |
4616 | return true; |
4617 | } | |
4618 | ||
4619 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
4620 | |
4621 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
4622 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
b8698a0f | 4623 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) |
8644a673 | 4624 | *dt = vect_external_def; |
ebfd146a IR |
4625 | else |
4626 | { | |
4627 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
4628 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
4629 | } | |
4630 | ||
4631 | if (*dt == vect_unknown_def_type) | |
4632 | { | |
4633 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4634 | fprintf (vect_dump, "Unsupported pattern."); | |
4635 | return false; | |
4636 | } | |
4637 | ||
4638 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4639 | fprintf (vect_dump, "type of def: %d.",*dt); | |
4640 | ||
4641 | switch (gimple_code (*def_stmt)) | |
4642 | { | |
4643 | case GIMPLE_PHI: | |
4644 | *def = gimple_phi_result (*def_stmt); | |
4645 | break; | |
4646 | ||
4647 | case GIMPLE_ASSIGN: | |
4648 | *def = gimple_assign_lhs (*def_stmt); | |
4649 | break; | |
4650 | ||
4651 | case GIMPLE_CALL: | |
4652 | *def = gimple_call_lhs (*def_stmt); | |
4653 | if (*def != NULL) | |
4654 | break; | |
4655 | /* FALLTHRU */ | |
4656 | default: | |
4657 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4658 | fprintf (vect_dump, "unsupported defining stmt: "); | |
4659 | return false; | |
4660 | } | |
4661 | ||
4662 | return true; | |
4663 | } | |
4664 | ||
b690cc0f RG |
4665 | /* Function vect_is_simple_use_1. |
4666 | ||
4667 | Same as vect_is_simple_use_1 but also determines the vector operand | |
4668 | type of OPERAND and stores it to *VECTYPE. If the definition of | |
4669 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
4670 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
4671 | is responsible to compute the best suited vector type for the | |
4672 | scalar operand. */ | |
4673 | ||
4674 | bool | |
4675 | vect_is_simple_use_1 (tree operand, loop_vec_info loop_vinfo, | |
4676 | bb_vec_info bb_vinfo, gimple *def_stmt, | |
4677 | tree *def, enum vect_def_type *dt, tree *vectype) | |
4678 | { | |
4679 | if (!vect_is_simple_use (operand, loop_vinfo, bb_vinfo, def_stmt, def, dt)) | |
4680 | return false; | |
4681 | ||
4682 | /* Now get a vector type if the def is internal, otherwise supply | |
4683 | NULL_TREE and leave it up to the caller to figure out a proper | |
4684 | type for the use stmt. */ | |
4685 | if (*dt == vect_internal_def | |
4686 | || *dt == vect_induction_def | |
4687 | || *dt == vect_reduction_def | |
4688 | || *dt == vect_double_reduction_def | |
4689 | || *dt == vect_nested_cycle) | |
4690 | { | |
4691 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
4692 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
4693 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
4694 | *vectype = STMT_VINFO_VECTYPE (stmt_info); | |
4695 | gcc_assert (*vectype != NULL_TREE); | |
4696 | } | |
4697 | else if (*dt == vect_uninitialized_def | |
4698 | || *dt == vect_constant_def | |
4699 | || *dt == vect_external_def) | |
4700 | *vectype = NULL_TREE; | |
4701 | else | |
4702 | gcc_unreachable (); | |
4703 | ||
4704 | return true; | |
4705 | } | |
4706 | ||
ebfd146a IR |
4707 | |
4708 | /* Function supportable_widening_operation | |
4709 | ||
b8698a0f L |
4710 | Check whether an operation represented by the code CODE is a |
4711 | widening operation that is supported by the target platform in | |
b690cc0f RG |
4712 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
4713 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 4714 | |
ebfd146a IR |
4715 | Widening operations we currently support are NOP (CONVERT), FLOAT |
4716 | and WIDEN_MULT. This function checks if these operations are supported | |
4717 | by the target platform either directly (via vector tree-codes), or via | |
4718 | target builtins. | |
4719 | ||
4720 | Output: | |
b8698a0f L |
4721 | - CODE1 and CODE2 are codes of vector operations to be used when |
4722 | vectorizing the operation, if available. | |
ebfd146a IR |
4723 | - DECL1 and DECL2 are decls of target builtin functions to be used |
4724 | when vectorizing the operation, if available. In this case, | |
b8698a0f | 4725 | CODE1 and CODE2 are CALL_EXPR. |
ebfd146a IR |
4726 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
4727 | case of multi-step conversion (like char->short->int - in that case | |
4728 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
4729 | - INTERM_TYPES contains the intermediate type required to perform the |
4730 | widening operation (short in the above example). */ | |
ebfd146a IR |
4731 | |
4732 | bool | |
b690cc0f RG |
4733 | supportable_widening_operation (enum tree_code code, gimple stmt, |
4734 | tree vectype_out, tree vectype_in, | |
ebfd146a IR |
4735 | tree *decl1, tree *decl2, |
4736 | enum tree_code *code1, enum tree_code *code2, | |
4737 | int *multi_step_cvt, | |
4738 | VEC (tree, heap) **interm_types) | |
4739 | { | |
4740 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4741 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4742 | struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info); | |
4743 | bool ordered_p; | |
4744 | enum machine_mode vec_mode; | |
81f40b79 | 4745 | enum insn_code icode1, icode2; |
ebfd146a | 4746 | optab optab1, optab2; |
b690cc0f RG |
4747 | tree vectype = vectype_in; |
4748 | tree wide_vectype = vectype_out; | |
ebfd146a IR |
4749 | enum tree_code c1, c2; |
4750 | ||
4751 | /* The result of a vectorized widening operation usually requires two vectors | |
b8698a0f L |
4752 | (because the widened results do not fit int one vector). The generated |
4753 | vector results would normally be expected to be generated in the same | |
ebfd146a IR |
4754 | order as in the original scalar computation, i.e. if 8 results are |
4755 | generated in each vector iteration, they are to be organized as follows: | |
b8698a0f | 4756 | vect1: [res1,res2,res3,res4], vect2: [res5,res6,res7,res8]. |
ebfd146a | 4757 | |
b8698a0f | 4758 | However, in the special case that the result of the widening operation is |
ebfd146a | 4759 | used in a reduction computation only, the order doesn't matter (because |
b8698a0f | 4760 | when vectorizing a reduction we change the order of the computation). |
ebfd146a IR |
4761 | Some targets can take advantage of this and generate more efficient code. |
4762 | For example, targets like Altivec, that support widen_mult using a sequence | |
4763 | of {mult_even,mult_odd} generate the following vectors: | |
4764 | vect1: [res1,res3,res5,res7], vect2: [res2,res4,res6,res8]. | |
4765 | ||
4766 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
b8698a0f L |
4767 | (each vectorized inner-loop iteration contributes to VF outer-loop |
4768 | iterations in parallel). We therefore don't allow to change the order | |
ebfd146a IR |
4769 | of the computation in the inner-loop during outer-loop vectorization. */ |
4770 | ||
4771 | if (STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
4772 | && !nested_in_vect_loop_p (vect_loop, stmt)) | |
4773 | ordered_p = false; | |
4774 | else | |
4775 | ordered_p = true; | |
4776 | ||
4777 | if (!ordered_p | |
4778 | && code == WIDEN_MULT_EXPR | |
4779 | && targetm.vectorize.builtin_mul_widen_even | |
4780 | && targetm.vectorize.builtin_mul_widen_even (vectype) | |
4781 | && targetm.vectorize.builtin_mul_widen_odd | |
4782 | && targetm.vectorize.builtin_mul_widen_odd (vectype)) | |
4783 | { | |
4784 | if (vect_print_dump_info (REPORT_DETAILS)) | |
4785 | fprintf (vect_dump, "Unordered widening operation detected."); | |
4786 | ||
4787 | *code1 = *code2 = CALL_EXPR; | |
4788 | *decl1 = targetm.vectorize.builtin_mul_widen_even (vectype); | |
4789 | *decl2 = targetm.vectorize.builtin_mul_widen_odd (vectype); | |
4790 | return true; | |
4791 | } | |
4792 | ||
4793 | switch (code) | |
4794 | { | |
4795 | case WIDEN_MULT_EXPR: | |
4796 | if (BYTES_BIG_ENDIAN) | |
4797 | { | |
4798 | c1 = VEC_WIDEN_MULT_HI_EXPR; | |
4799 | c2 = VEC_WIDEN_MULT_LO_EXPR; | |
4800 | } | |
4801 | else | |
4802 | { | |
4803 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
4804 | c1 = VEC_WIDEN_MULT_LO_EXPR; | |
4805 | } | |
4806 | break; | |
4807 | ||
4808 | CASE_CONVERT: | |
4809 | if (BYTES_BIG_ENDIAN) | |
4810 | { | |
4811 | c1 = VEC_UNPACK_HI_EXPR; | |
4812 | c2 = VEC_UNPACK_LO_EXPR; | |
4813 | } | |
4814 | else | |
4815 | { | |
4816 | c2 = VEC_UNPACK_HI_EXPR; | |
4817 | c1 = VEC_UNPACK_LO_EXPR; | |
4818 | } | |
4819 | break; | |
4820 | ||
4821 | case FLOAT_EXPR: | |
4822 | if (BYTES_BIG_ENDIAN) | |
4823 | { | |
4824 | c1 = VEC_UNPACK_FLOAT_HI_EXPR; | |
4825 | c2 = VEC_UNPACK_FLOAT_LO_EXPR; | |
4826 | } | |
4827 | else | |
4828 | { | |
4829 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
4830 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; | |
4831 | } | |
4832 | break; | |
4833 | ||
4834 | case FIX_TRUNC_EXPR: | |
4835 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
4836 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
4837 | computing the operation. */ | |
4838 | return false; | |
4839 | ||
4840 | default: | |
4841 | gcc_unreachable (); | |
4842 | } | |
4843 | ||
4844 | if (code == FIX_TRUNC_EXPR) | |
4845 | { | |
4846 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
4847 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
4848 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
4849 | } |
4850 | else | |
4851 | { | |
4852 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
4853 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
4854 | } | |
4855 | ||
4856 | if (!optab1 || !optab2) | |
4857 | return false; | |
4858 | ||
4859 | vec_mode = TYPE_MODE (vectype); | |
4860 | if ((icode1 = optab_handler (optab1, vec_mode)->insn_code) == CODE_FOR_nothing | |
4861 | || (icode2 = optab_handler (optab2, vec_mode)->insn_code) | |
4862 | == CODE_FOR_nothing) | |
4863 | return false; | |
4864 | ||
b8698a0f | 4865 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a IR |
4866 | types. */ |
4867 | if (insn_data[icode1].operand[0].mode != TYPE_MODE (wide_vectype) | |
4868 | || insn_data[icode2].operand[0].mode != TYPE_MODE (wide_vectype)) | |
4869 | { | |
4870 | int i; | |
4871 | tree prev_type = vectype, intermediate_type; | |
4872 | enum machine_mode intermediate_mode, prev_mode = vec_mode; | |
4873 | optab optab3, optab4; | |
4874 | ||
4875 | if (!CONVERT_EXPR_CODE_P (code)) | |
4876 | return false; | |
b8698a0f | 4877 | |
ebfd146a IR |
4878 | *code1 = c1; |
4879 | *code2 = c2; | |
b8698a0f | 4880 | |
ebfd146a IR |
4881 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
4882 | intermediate steps in promotion sequence. We try MAX_INTERM_CVT_STEPS | |
4883 | to get to NARROW_VECTYPE, and fail if we do not. */ | |
4884 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); | |
4885 | for (i = 0; i < 3; i++) | |
4886 | { | |
4887 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
4888 | intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode, | |
4889 | TYPE_UNSIGNED (prev_type)); | |
4890 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
4891 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
4892 | ||
4893 | if (!optab3 || !optab4 | |
4894 | || (icode1 = optab1->handlers[(int) prev_mode].insn_code) | |
4895 | == CODE_FOR_nothing | |
4896 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
4897 | || (icode2 = optab2->handlers[(int) prev_mode].insn_code) | |
4898 | == CODE_FOR_nothing | |
4899 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
b8698a0f | 4900 | || (icode1 = optab3->handlers[(int) intermediate_mode].insn_code) |
ebfd146a IR |
4901 | == CODE_FOR_nothing |
4902 | || (icode2 = optab4->handlers[(int) intermediate_mode].insn_code) | |
4903 | == CODE_FOR_nothing) | |
4904 | return false; | |
4905 | ||
4906 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
4907 | (*multi_step_cvt)++; | |
4908 | ||
4909 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
4910 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
4911 | return true; | |
4912 | ||
4913 | prev_type = intermediate_type; | |
4914 | prev_mode = intermediate_mode; | |
4915 | } | |
4916 | ||
4917 | return false; | |
4918 | } | |
4919 | ||
4920 | *code1 = c1; | |
4921 | *code2 = c2; | |
4922 | return true; | |
4923 | } | |
4924 | ||
4925 | ||
4926 | /* Function supportable_narrowing_operation | |
4927 | ||
b8698a0f L |
4928 | Check whether an operation represented by the code CODE is a |
4929 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
4930 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
4931 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 4932 | |
ebfd146a IR |
4933 | Narrowing operations we currently support are NOP (CONVERT) and |
4934 | FIX_TRUNC. This function checks if these operations are supported by | |
4935 | the target platform directly via vector tree-codes. | |
4936 | ||
4937 | Output: | |
b8698a0f L |
4938 | - CODE1 is the code of a vector operation to be used when |
4939 | vectorizing the operation, if available. | |
ebfd146a IR |
4940 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
4941 | case of multi-step conversion (like int->short->char - in that case | |
4942 | MULTI_STEP_CVT will be 1). | |
4943 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 4944 | narrowing operation (short in the above example). */ |
ebfd146a IR |
4945 | |
4946 | bool | |
4947 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 4948 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
4949 | enum tree_code *code1, int *multi_step_cvt, |
4950 | VEC (tree, heap) **interm_types) | |
4951 | { | |
4952 | enum machine_mode vec_mode; | |
4953 | enum insn_code icode1; | |
4954 | optab optab1, interm_optab; | |
b690cc0f RG |
4955 | tree vectype = vectype_in; |
4956 | tree narrow_vectype = vectype_out; | |
ebfd146a IR |
4957 | enum tree_code c1; |
4958 | tree intermediate_type, prev_type; | |
4959 | int i; | |
4960 | ||
4961 | switch (code) | |
4962 | { | |
4963 | CASE_CONVERT: | |
4964 | c1 = VEC_PACK_TRUNC_EXPR; | |
4965 | break; | |
4966 | ||
4967 | case FIX_TRUNC_EXPR: | |
4968 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
4969 | break; | |
4970 | ||
4971 | case FLOAT_EXPR: | |
4972 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
4973 | tree code and optabs used for computing the operation. */ | |
4974 | return false; | |
4975 | ||
4976 | default: | |
4977 | gcc_unreachable (); | |
4978 | } | |
4979 | ||
4980 | if (code == FIX_TRUNC_EXPR) | |
4981 | /* The signedness is determined from output operand. */ | |
b690cc0f | 4982 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
4983 | else |
4984 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
4985 | ||
4986 | if (!optab1) | |
4987 | return false; | |
4988 | ||
4989 | vec_mode = TYPE_MODE (vectype); | |
b8698a0f | 4990 | if ((icode1 = optab_handler (optab1, vec_mode)->insn_code) |
ebfd146a IR |
4991 | == CODE_FOR_nothing) |
4992 | return false; | |
4993 | ||
4994 | /* Check if it's a multi-step conversion that can be done using intermediate | |
4995 | types. */ | |
4996 | if (insn_data[icode1].operand[0].mode != TYPE_MODE (narrow_vectype)) | |
4997 | { | |
4998 | enum machine_mode intermediate_mode, prev_mode = vec_mode; | |
4999 | ||
5000 | *code1 = c1; | |
5001 | prev_type = vectype; | |
5002 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS | |
5003 | intermediate steps in promotion sequence. We try MAX_INTERM_CVT_STEPS | |
5004 | to get to NARROW_VECTYPE, and fail if we do not. */ | |
5005 | *interm_types = VEC_alloc (tree, heap, MAX_INTERM_CVT_STEPS); | |
5006 | for (i = 0; i < 3; i++) | |
5007 | { | |
5008 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
5009 | intermediate_type = lang_hooks.types.type_for_mode (intermediate_mode, | |
5010 | TYPE_UNSIGNED (prev_type)); | |
b8698a0f | 5011 | interm_optab = optab_for_tree_code (c1, intermediate_type, |
ebfd146a | 5012 | optab_default); |
b8698a0f | 5013 | if (!interm_optab |
ebfd146a IR |
5014 | || (icode1 = optab1->handlers[(int) prev_mode].insn_code) |
5015 | == CODE_FOR_nothing | |
5016 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
b8698a0f | 5017 | || (icode1 |
ebfd146a IR |
5018 | = interm_optab->handlers[(int) intermediate_mode].insn_code) |
5019 | == CODE_FOR_nothing) | |
5020 | return false; | |
5021 | ||
5022 | VEC_quick_push (tree, *interm_types, intermediate_type); | |
5023 | (*multi_step_cvt)++; | |
5024 | ||
5025 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5026 | return true; | |
5027 | ||
5028 | prev_type = intermediate_type; | |
5029 | prev_mode = intermediate_mode; | |
5030 | } | |
5031 | ||
5032 | return false; | |
5033 | } | |
5034 | ||
5035 | *code1 = c1; | |
5036 | return true; | |
5037 | } |